{
  "article": {
    "assignment_desk_version": "v2",
    "assignment_rank": 12,
    "canonical_story_id": "a6e0808c43a6a023",
    "canonical_url": "/news/story/hc_2fc7469e80ee3da77a3b3a81/",
    "cluster_manifest_ids": [
      "1790097172664828006",
      "1790093784543303523",
      "1790093784543444754",
      "1790093784543011548",
      "1790093784543906695",
      "1790093784543625410"
    ],
    "cluster_signature": "title:new datasets and pipelines accelerate robotic manipulation and surgical training",
    "cluster_stream_ids": [
      "17779468859058016",
      "17780442148971635",
      "17894412312039768"
    ],
    "collapsed_duplicate_domains": [],
    "computed_evidence_counts": {
      "claims": 18,
      "earliest_published": "2026-09-22T00:00:00+00:00",
      "latest_published": "2026-09-22T00:00:00+00:00",
      "manifests": 6,
      "sources": 2,
      "span_hours": 0.0
    },
    "cross_domain_evidence": false,
    "dedupe_reason": "canonical public story",
    "discovery": {
      "excluded_manifest_count": 0,
      "excluded_manifest_ids": [],
      "manifest_fetch_count": 40,
      "manifest_fetch_strategy": "seed_clusters_then_high_significance",
      "mcp_window": {
        "published_date_from": "2026-09-22",
        "published_date_to": "2026-09-23"
      },
      "preflight_count": 462,
      "read_cap": 40,
      "seed_manifest_fetches": [
        {
          "excluded": 0,
          "kept": 40,
          "label": "arxiv",
          "returned": 40,
          "source_channel": "arxiv",
          "stream_id": "17779468859058016"
        }
      ],
      "seed_preflights": [
        {
          "baseline_daily": 138.71428571428572,
          "count_24h": 51,
          "event_count": 0,
          "high_count": 34,
          "kind": "stream_cluster",
          "label": "arxiv",
          "mcp_count": 229,
          "score": 153.37,
          "source_channel": "arxiv",
          "spike_ratio": 0.37,
          "stream_id": "17779468859058016"
        }
      ],
      "strict_filter": "manifest timestamp within trailing 12h"
    },
    "edition_date": "2026-09-23",
    "edition_slot": "00",
    "evidence_manifest_ids": [
      "1790093784543011548",
      "1790093784543303523",
      "1790093784543444754",
      "1790093784543625410",
      "1790093784543906695",
      "1790097172664828006"
    ],
    "evidence_stream_ids": [
      "17779468859058016",
      "17780442148971635",
      "17894412312039768",
      "17894502471992576"
    ],
    "excluded_manifest_ids": [],
    "featured_claims": [
      {
        "claim_id": "1790095196739357477",
        "claim_ref": "015aa72a3568db05db64c126463d81c2659a1b31",
        "headline": "HuRo: Robotizing Human Videos for Scalable VLA Pretraining",
        "manifest_id": "1790093784543444754",
        "source_channel": "arXiv - Official Multimodal Document AI",
        "source_url": "https://arxiv.org/pdf/2609.10706v4",
        "text": "Pretraining a VLA policy on HuRo data improves task completion from 51.5% to 80.3%."
      },
      {
        "claim_id": "1790097763131447727",
        "claim_ref": "c564a6e56771d1f4bcc5a887d7e349b614ff7f02",
        "headline": "HumynexSurg-1: A Curated Expert Liposuction Dataset",
        "manifest_id": "1790097172664828006",
        "source_channel": "arXiv Frontier AI Methods Select",
        "source_url": "https://arxiv.org/pdf/2609.23885v1",
        "text": "NVIDIA Isaac GR00T N1.7 can be fine-tuned on the HumynexSurg-1 dataset in under one hour per run."
      }
    ],
    "home_domain": "engineering-technology",
    "kind": "domain_digest",
    "lead_citations": [
      {
        "claim_id": "1790095196739357477",
        "kind": "claim",
        "manifest_id": "1790093784543444754"
      },
      {
        "claim_id": "1790097763131447727",
        "kind": "claim",
        "manifest_id": "1790097172664828006"
      }
    ],
    "meta_brief": {
      "corroborated_takeaways": 0,
      "facts": [
        {
          "disputed": false,
          "label": "Episodes",
          "readings": [
            {
              "manifest_id": "1790097172664828006",
              "value": "14"
            }
          ],
          "sources": 1
        },
        {
          "disputed": false,
          "label": "Duration",
          "readings": [
            {
              "manifest_id": "1790097172664828006",
              "value": "35.6"
            }
          ],
          "sources": 1
        },
        {
          "disputed": false,
          "label": "Utterances",
          "readings": [
            {
              "manifest_id": "1790097172664828006",
              "value": "356"
            }
          ],
          "sources": 1
        },
        {
          "disputed": false,
          "label": "Accuracy",
          "readings": [
            {
              "manifest_id": "1790093784543303523",
              "value": "95"
            }
          ],
          "sources": 1
        },
        {
          "disputed": false,
          "label": "Improvement over baselines",
          "readings": [
            {
              "manifest_id": "1790093784543303523",
              "value": "6"
            }
          ],
          "sources": 1
        },
        {
          "disputed": false,
          "label": "HuRo dataset size",
          "readings": [
            {
              "manifest_id": "1790093784543444754",
              "value": "630K"
            }
          ],
          "sources": 1
        },
        {
          "disputed": false,
          "label": "Processed frames",
          "readings": [
            {
              "manifest_id": "1790093784543444754",
              "value": "142M"
            }
          ],
          "sources": 1
        },
        {
          "disputed": false,
          "label": "Task completion improvement",
          "readings": [
            {
              "manifest_id": "1790093784543444754",
              "value": "51.5 to 80.3"
            }
          ],
          "sources": 1
        },
        {
          "disputed": false,
          "label": "OOD completion improvement",
          "readings": [
            {
              "manifest_id": "1790093784543444754",
              "value": "34.9 to 72.2"
            }
          ],
          "sources": 1
        },
        {
          "disputed": false,
          "label": "images",
          "readings": [
            {
              "manifest_id": "1790093784543625410",
              "value": "446"
            }
          ],
          "sources": 1
        },
        {
          "disputed": false,
          "label": "pathologists",
          "readings": [
            {
              "manifest_id": "1790093784543625410",
              "value": "4"
            }
          ],
          "sources": 1
        }
      ],
      "manifest_count": 6,
      "open_questions": [
        {
          "manifest_ids": [
            "1790097172664828006"
          ],
          "sources": 1,
          "text": "No public URL provided for the dataset in the abstract"
        },
        {
          "manifest_ids": [
            "1790097172664828006"
          ],
          "sources": 1,
          "text": "Requires further clinical validation"
        },
        {
          "manifest_ids": [
            "1790093784543303523"
          ],
          "sources": 1,
          "text": "No public code or dataset repository link provided in the abstract."
        },
        {
          "manifest_ids": [
            "1790093784543444754"
          ],
          "sources": 1,
          "text": "No explicit mention of open-source code release in the abstract, only a project website."
        },
        {
          "manifest_ids": [
            "1790093784543011548"
          ],
          "sources": 1,
          "text": "No clear performance advantage for AGM over RM across general object discovery tasks."
        },
        {
          "manifest_ids": [
            "1790093784543906695"
          ],
          "sources": 1,
          "text": "Code, prompts, and validator scripts are not yet released."
        }
      ],
      "source_tldrs": [
        {
          "manifest_id": "1790097172664828006",
          "text": "A new multimodal dataset for robotic liposuction enables fine-tuning of foundation models for complex, force-sensitive surgical tasks."
        },
        {
          "manifest_id": "1790093784543303523",
          "text": "A Multimodal Fusion Transformer improves pedestrian distraction detection by 6% over baselines by integrating skeletal dynamics and visual appearance."
        },
        {
          "manifest_id": "1790093784543444754",
          "text": "A pipeline that converts human videos into robot-aligned training data, significantly boosting VLA policy performance and OOD robustness."
        },
        {
          "manifest_id": "1790093784543011548",
          "text": "Attention-guided masking provides no consistent performance benefit over random masking for object discovery in object-centric learning."
        },
        {
          "manifest_id": "1790093784543906695",
          "text": "A verifiable, cost-efficient framework for autonomous driving that uses a structured Perception-Evaluation-Revision chain to reduce reliance on expensive frontier model supervision."
        },
        {
          "manifest_id": "1790093784543625410",
          "text": "CytoCrowd provides a multi-annotator cytology dataset featuring both conflicting expert labels and a senior gold standard to improve medical AI robustness."
        }
      ],
      "stakes": [
        {
          "manifest_id": "1790097172664828006",
          "text": "Accelerates the development of autonomous surgical robotics by providing high-quality, expert-demonstration data for non-visual manipulation."
        },
        {
          "manifest_id": "1790093784543303523",
          "text": "Enhances safety-critical perception systems for autonomous vehicle operators and AI developers."
        },
        {
          "manifest_id": "1790093784543444754",
          "text": "Essential for robotics researchers and AI builders looking to scale VLA training beyond limited real-robot datasets."
        },
        {
          "manifest_id": "1790093784543011548",
          "text": "Informs researchers and AI engineers on effective self-supervision strategies for object-centric representation learning, preventing wasted compute on ineffective architectural modifications."
        },
        {
          "manifest_id": "1790093784543906695",
          "text": "Enables AI operators and developers to scale high-quality training data for autonomous systems while significantly reducing operational inference costs."
        },
        {
          "manifest_id": "1790093784543625410",
          "text": "Essential for developers and researchers building medical diagnostic AI that must account for subjective expert disagreement and high-stakes labeling variance."
        }
      ],
      "takeaways": [
        {
          "manifest_ids": [
            "1790097172664828006"
          ],
          "sources": 1,
          "text": "Demonstration that NVIDIA Isaac GR00T N1.7 can be fine-tuned on this dataset in under an hour per run."
        },
        {
          "manifest_ids": [
            "1790097172664828006"
          ],
          "sources": 1,
          "text": "Scaling analysis indicates that performance improves with each additional session, suggesting a clear path for data-driven surgical robotics."
        },
        {
          "manifest_ids": [
            "1790097172664828006"
          ],
          "sources": 1,
          "text": "Integrate HumynexSurg-1 into robotic manipulation pipelines to evaluate foundation model performance on high-precision, non-visual surgical tasks."
        },
        {
          "manifest_ids": [
            "1790093784543303523"
          ],
          "sources": 1,
          "text": "Achieved 95% accuracy, outperforming six baseline models by 6%."
        },
        {
          "manifest_ids": [
            "1790093784543303523"
          ],
          "sources": 1,
          "text": "Demonstrates the efficacy of combining skeletal dynamics with visual appearance for safety-critical computer vision tasks."
        },
        {
          "manifest_ids": [
            "1790093784543303523"
          ],
          "sources": 1,
          "text": "Integrate skeletal-visual fusion architectures into autonomous vehicle perception stacks to improve pedestrian risk assessment."
        },
        {
          "manifest_ids": [
            "1790093784543303523"
          ],
          "sources": 1,
          "text": "Evaluate MFT performance against existing baseline models in edge-case distraction scenarios."
        },
        {
          "manifest_ids": [
            "1790093784543444754"
          ],
          "sources": 1,
          "text": "HuRo dataset contains 630K episodes and 142M frames from five human-video sources."
        }
      ]
    },
    "newsworthiness_score": {
      "audience_fit": 0.55,
      "breadth": 0.5125,
      "domain_priority": 0.98,
      "materiality": 0.9,
      "novelty": 1.0,
      "source_strength": 0.75,
      "total": 0.78725
    },
    "paragraphs": [],
    "phase": "curated_synthesis",
    "primary_home_domain": "engineering-technology",
    "prompt_version": "news_editorial_v2",
    "schema_version": 1,
    "secondary_home_domains": [],
    "seed_candidates": [
      {
        "baseline_daily": 138.71428571428572,
        "count_24h": 51,
        "event_count": 0,
        "high_count": 34,
        "kind": "stream_cluster",
        "label": "arxiv",
        "score": 153.37,
        "source_channel": "arxiv",
        "spike_ratio": 0.37,
        "stream_id": "17779468859058016"
      },
      {
        "baseline_daily": 6.5,
        "count_24h": 26,
        "event_count": 0,
        "high_count": 19,
        "kind": "stream_cluster",
        "label": "arxiv-frontier-methods-select",
        "score": 87.0,
        "source_channel": null,
        "spike_ratio": 4.0,
        "stream_id": null
      },
      {
        "baseline_daily": 16.428571428571427,
        "count_24h": 26,
        "event_count": 0,
        "high_count": 19,
        "kind": "stream_cluster",
        "label": "chemrxiv",
        "score": 84.58,
        "source_channel": null,
        "spike_ratio": 1.58,
        "stream_id": null
      },
      {
        "baseline_daily": 7.571428571428571,
        "count_24h": 25,
        "event_count": 0,
        "high_count": 15,
        "kind": "stream_cluster",
        "label": "arxiv-multimodal-document-ai",
        "score": 73.3,
        "source_channel": null,
        "spike_ratio": 3.3,
        "stream_id": null
      },
      {
        "baseline_daily": 0.0,
        "count_24h": 17,
        "event_count": 0,
        "high_count": 13,
        "kind": "stream_cluster",
        "label": "unknown-stream",
        "score": 66.0,
        "source_channel": "tesla-blog",
        "spike_ratio": 17.0,
        "stream_id": null
      },
      {
        "baseline_daily": 1.5714285714285714,
        "count_24h": 9,
        "event_count": 0,
        "high_count": 6,
        "kind": "stream_cluster",
        "label": "cisa-cybersecurity-advisories",
        "score": 32.73,
        "source_channel": null,
        "spike_ratio": 5.73,
        "stream_id": null
      },
      {
        "baseline_daily": 11.5,
        "count_24h": 9,
        "event_count": 0,
        "high_count": 7,
        "kind": "stream_cluster",
        "label": "biorxiv",
        "score": 30.78,
        "source_channel": null,
        "spike_ratio": 0.78,
        "stream_id": null
      },
      {
        "baseline_daily": 0.07142857142857142,
        "count_24h": 1,
        "event_count": 0,
        "high_count": 1,
        "kind": "stream_cluster",
        "label": "biorxiv-cell-molecular-mechanisms",
        "score": 14.0,
        "source_channel": null,
        "spike_ratio": 10.0,
        "stream_id": null
      },
      {
        "baseline_daily": 3.2857142857142856,
        "count_24h": 6,
        "event_count": 0,
        "high_count": 2,
        "kind": "stream_cluster",
        "label": "medrxiv",
        "score": 13.83,
        "source_channel": null,
        "spike_ratio": 1.83,
        "stream_id": null
      },
      {
        "baseline_daily": 0.21428571428571427,
        "count_24h": 1,
        "event_count": 0,
        "high_count": 1,
        "kind": "stream_cluster",
        "label": "trm-labs-security-research",
        "score": 8.67,
        "source_channel": "trm-labs-security-research",
        "spike_ratio": 4.67,
        "stream_id": "17870941650738828"
      },
      {
        "baseline_daily": 0.2857142857142857,
        "count_24h": 1,
        "event_count": 0,
        "high_count": 1,
        "kind": "stream_cluster",
        "label": "bitsight",
        "score": 7.5,
        "source_channel": "bitsight",
        "spike_ratio": 3.5,
        "stream_id": "17870549302903372"
      },
      {
        "baseline_daily": 0.35714285714285715,
        "count_24h": 1,
        "event_count": 0,
        "high_count": 1,
        "kind": "stream_cluster",
        "label": "anomali",
        "score": 6.8,
        "source_channel": "anomali",
        "spike_ratio": 2.8,
        "stream_id": "17863880757312444"
      }
    ],
    "slug": "00-12-robotic-multimodal-pretraining-datasets",
    "source_manifests": [
      {
        "brief": {
          "actionable_takeaways": [
            "Integrate HumynexSurg-1 into robotic manipulation pipelines to evaluate foundation model performance on high-precision, non-visual surgical tasks."
          ],
          "facts": [
            {
              "label": "Episodes",
              "value": "14"
            },
            {
              "label": "Duration",
              "value": "35.6"
            },
            {
              "label": "Utterances",
              "value": "356"
            }
          ],
          "key_insights": [
            "Demonstration that NVIDIA Isaac GR00T N1.7 can be fine-tuned on this dataset in under an hour per run.",
            "Scaling analysis indicates that performance improves with each additional session, suggesting a clear path for data-driven surgical robotics."
          ],
          "tldr": "A new multimodal dataset for robotic liposuction enables fine-tuning of foundation models for complex, force-sensitive surgical tasks.",
          "unresolved": [
            "No public URL provided for the dataset in the abstract",
            "Requires further clinical validation"
          ],
          "why_it_matters": "Accelerates the development of autonomous surgical robotics by providing high-quality, expert-demonstration data for non-visual manipulation."
        },
        "claim_count": 3,
        "claims": [
          {
            "claim_id": "1790097763129214790",
            "claim_ref": "30575459e5c669ba4418cdbcd859d8f7b9edfe39",
            "snapshot": "{\"claim_id\":\"1790097763129214790\",\"claim_text\":\"HumynexSurg-1 is the first curated expert liposuction dataset featuring synchronized force/torque, RGB-D video, and expert narration.\",\"claim_type\":\"statement\",\"confidence\":\"measured\",\"entities\":[{\"name\":\"Humynex Robotics\",\"role\":\"subject\",\"type\":\"organization\"}],\"evidence\":\"paraphrase\",\"featured\":true,\"key_point_index\":null,\"quote\":null,\"signal\":\"Monitor for benchmarking robotic surgical manipulation and evaluating multimodal foundation model performance in medical contexts.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "HumynexSurg-1 is the first curated expert liposuction dataset featuring synchronized force/torque, RGB-D video, and expert narration."
          },
          {
            "claim_id": "1790097763131447727",
            "claim_ref": "c564a6e56771d1f4bcc5a887d7e349b614ff7f02",
            "snapshot": "{\"claim_id\":\"1790097763131447727\",\"claim_text\":\"NVIDIA Isaac GR00T N1.7 can be fine-tuned on the HumynexSurg-1 dataset in under one hour per run.\",\"claim_type\":\"data\",\"confidence\":\"measured\",\"entities\":[{\"name\":\"NVIDIA\",\"role\":\"mentioned\",\"tag_id\":\"17723038993599085\",\"type\":\"organization\"}],\"evidence\":\"derived\",\"featured\":true,\"key_point_index\":null,\"quote\":null,\"signal\":\"Evaluate model training efficiency for specialized robotic tasks; use as a baseline for fine-tuning performance on surgical datasets.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "NVIDIA Isaac GR00T N1.7 can be fine-tuned on the HumynexSurg-1 dataset in under one hour per run."
          },
          {
            "claim_id": "1790097763133081034",
            "claim_ref": "7b96f807ce2bff8135ae3ea15edf535e49cebeef",
            "snapshot": "{\"claim_id\":\"1790097763133081034\",\"claim_text\":\"Every new session added to the training set lowers the error on unseen sessions.\",\"claim_type\":\"analysis\",\"confidence\":\"measured\",\"entities\":[{\"name\":\"Humynex Robotics\",\"role\":\"subject\",\"type\":\"organization\"}],\"evidence\":\"observed\",\"featured\":true,\"key_point_index\":null,\"quote\":null,\"signal\":\"Use to inform data collection strategies and scaling laws for robotic foundation models in high-stakes physical domains.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "Every new session added to the training set lowers the error on unseen sessions."
          }
        ],
        "entities": [],
        "headline": "HumynexSurg-1: A Curated Expert Liposuction Dataset",
        "home_domain": null,
        "manifest_id": "1790097172664828006",
        "published_at": "2026-09-22",
        "significance": "high",
        "source_channel": "arXiv Frontier AI Methods Select",
        "source_name": "arxiv-frontier-methods-select",
        "source_url": "https://arxiv.org/pdf/2609.23885v1",
        "stream_id": null,
        "stream_ids": [
          "17779468859058016",
          "17780442148971635",
          "17894412312039768"
        ],
        "summary": "HumynexSurg-1 addresses the critical shortage of surgical demonstration data for robot foundation models by providing a high-fidelity, multimodal dataset for liposuction. It includes synchronized force/torque, RGB-D video, and expert narration, enabling training for complex, feel-based surgical tasks. This release is significant for AI developers and robotics researchers aiming to bridge the gap between general manipulation and specialized medical procedures.",
        "tags": [
          "arXiv",
          "Multimodal AI"
        ]
      },
      {
        "brief": {
          "actionable_takeaways": [
            "Integrate skeletal-visual fusion architectures into autonomous vehicle perception stacks to improve pedestrian risk assessment.",
            "Evaluate MFT performance against existing baseline models in edge-case distraction scenarios."
          ],
          "facts": [
            {
              "label": "Accuracy",
              "value": "95"
            },
            {
              "label": "Improvement over baselines",
              "value": "6"
            }
          ],
          "key_insights": [
            "Achieved 95% accuracy, outperforming six baseline models by 6%.",
            "Demonstrates the efficacy of combining skeletal dynamics with visual appearance for safety-critical computer vision tasks."
          ],
          "tldr": "A Multimodal Fusion Transformer improves pedestrian distraction detection by 6% over baselines by integrating skeletal dynamics and visual appearance.",
          "unresolved": [
            "No public code or dataset repository link provided in the abstract."
          ],
          "why_it_matters": "Enhances safety-critical perception systems for autonomous vehicle operators and AI developers."
        },
        "claim_count": 3,
        "claims": [
          {
            "claim_id": "1790095228663526349",
            "claim_ref": "2a37913d0c705ec150a39f092c2a098988c07084",
            "snapshot": "{\"claim_id\":\"1790095228663526349\",\"claim_text\":\"MFT achieves 95% accuracy in detecting phone-induced pedestrian distraction.\",\"claim_type\":\"data\",\"confidence\":\"measured\",\"entities\":[{\"name\":\"multimodal_ai\",\"role\":\"mentioned\",\"tag_id\":\"17730948119041167\",\"type\":\"topic\"}],\"evidence\":\"derived\",\"featured\":true,\"key_point_index\":null,\"quote\":null,\"signal\":\"Monitor this benchmark for autonomous vehicle perception stacks; compare against existing vision-only distraction detection models.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "MFT achieves 95% accuracy in detecting phone-induced pedestrian distraction."
          },
          {
            "claim_id": "1790095228680439855",
            "claim_ref": "5f3cc65da61c4cb1f15bb4f755a7924ac4da1261",
            "snapshot": "{\"claim_id\":\"1790095228680439855\",\"claim_text\":\"MFT outperforms six baseline approaches by 6% in pedestrian distraction detection.\",\"claim_type\":\"data\",\"confidence\":\"measured\",\"entities\":[{\"name\":\"multimodal_ai\",\"role\":\"mentioned\",\"tag_id\":\"17730948119041167\",\"type\":\"topic\"}],\"evidence\":\"derived\",\"featured\":true,\"key_point_index\":null,\"quote\":null,\"signal\":\"Use as a performance benchmark for evaluating new multimodal fusion architectures in safety-critical applications.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "MFT outperforms six baseline approaches by 6% in pedestrian distraction detection."
          },
          {
            "claim_id": "1790095228687149988",
            "claim_ref": "26ecf5fc1a3847ea1e4ee34dd1bc1703e5a1aa44",
            "snapshot": "{\"claim_id\":\"1790095228687149988\",\"claim_text\":\"Cross-modal attention effectively captures inter-modal dependencies between skeletal dynamics and visual appearance.\",\"claim_type\":\"statement\",\"confidence\":\"stated\",\"entities\":[{\"name\":\"multimodal_ai\",\"role\":\"mentioned\",\"tag_id\":\"17730948119041167\",\"type\":\"topic\"}],\"evidence\":\"paraphrase\",\"featured\":true,\"key_point_index\":null,\"quote\":null,\"signal\":\"Investigate the use of cross-modal attention modules for improving fusion in other multimodal perception tasks.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "Cross-modal attention effectively captures inter-modal dependencies between skeletal dynamics and visual appearance."
          }
        ],
        "entities": [],
        "headline": "Detecting Phone-Induced Pedestrian Distraction via a Multimodal Fusion Transformer",
        "home_domain": null,
        "manifest_id": "1790093784543303523",
        "published_at": "2026-09-22",
        "significance": "medium",
        "source_channel": "arXiv - Official Multimodal Document AI",
        "source_name": "arxiv-multimodal-document-ai",
        "source_url": "https://arxiv.org/pdf/2609.23507v1",
        "stream_id": null,
        "stream_ids": [
          "17779468859058016",
          "17894502471992576"
        ],
        "summary": "This research introduces a Multimodal Fusion Transformer (MFT) designed to identify distracted pedestrians by fusing skeletal pose data with visual appearance features. The model utilizes cross-modal and temporal attention mechanisms to improve detection accuracy, reaching 95% on a custom dataset. This approach offers significant utility for autonomous vehicle developers aiming to enhance real-time risk assessment and pedestrian safety.",
        "tags": [
          "arXiv",
          "Multimodal AI"
        ]
      },
      {
        "brief": {
          "actionable_takeaways": [
            "Integrate HuRo-style video-to-robot data conversion pipelines to augment VLA training sets.",
            "Prioritize visual robotization and retargeted action pretraining to enhance OOD robustness in manipulation tasks."
          ],
          "facts": [
            {
              "label": "HuRo dataset size",
              "value": "630K"
            },
            {
              "label": "Processed frames",
              "value": "142M"
            },
            {
              "label": "Task completion improvement",
              "value": "51.5 to 80.3"
            },
            {
              "label": "OOD completion improvement",
              "value": "34.9 to 72.2"
            }
          ],
          "key_insights": [
            "HuRo dataset contains 630K episodes and 142M frames from five human-video sources.",
            "Pretraining VLA policies on HuRo data improved task completion from 51.5% to 80.3% and OOD completion from 34.9% to 72.2%."
          ],
          "tldr": "A pipeline that converts human videos into robot-aligned training data, significantly boosting VLA policy performance and OOD robustness.",
          "unresolved": [
            "No explicit mention of open-source code release in the abstract, only a project website."
          ],
          "why_it_matters": "Essential for robotics researchers and AI builders looking to scale VLA training beyond limited real-robot datasets."
        },
        "claim_count": 4,
        "claims": [
          {
            "claim_id": "1790095196719128958",
            "claim_ref": "72f40efb17977f613638929bd4c9017886681b56",
            "snapshot": "{\"claim_id\":\"1790095196719128958\",\"claim_text\":\"A robotization pipeline can convert heterogeneous human videos into robot-aligned observations and action trajectories.\",\"claim_type\":\"statement\",\"confidence\":\"measured\",\"entities\":[{\"name\":\"HuRo\",\"role\":\"subject\",\"tag_id\":\"17891279667987836\",\"type\":\"topic\"}],\"evidence\":\"derived\",\"featured\":true,\"key_point_index\":1,\"quote\":null,\"signal\":\"Monitor this pipeline as a method for scaling VLA training data without requiring additional real-robot data collection.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "A robotization pipeline can convert heterogeneous human videos into robot-aligned observations and action trajectories."
          },
          {
            "claim_id": "1790095196735321611",
            "claim_ref": "87d152aef77a66a78d04c9e3f79a7df25146af96",
            "snapshot": "{\"claim_id\":\"1790095196735321611\",\"claim_text\":\"The HuRo dataset comprises 630K robotized episodes and 142M processed frames.\",\"claim_type\":\"data\",\"confidence\":\"stated\",\"entities\":[{\"name\":\"HuRo\",\"role\":\"subject\",\"tag_id\":\"17891279667987836\",\"type\":\"topic\"}],\"evidence\":\"derived\",\"featured\":true,\"key_point_index\":2,\"quote\":null,\"signal\":\"Evaluate the scale and diversity of this dataset for potential integration into VLA pretraining workflows.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "The HuRo dataset comprises 630K robotized episodes and 142M processed frames."
          },
          {
            "claim_id": "1790095196739357477",
            "claim_ref": "015aa72a3568db05db64c126463d81c2659a1b31",
            "snapshot": "{\"claim_id\":\"1790095196739357477\",\"claim_text\":\"Pretraining a VLA policy on HuRo data improves task completion from 51.5% to 80.3%.\",\"claim_type\":\"data\",\"confidence\":\"measured\",\"entities\":[{\"name\":\"HuRo\",\"role\":\"subject\",\"tag_id\":\"17891279667987836\",\"type\":\"topic\"}],\"evidence\":\"derived\",\"featured\":true,\"key_point_index\":3,\"quote\":null,\"signal\":\"Benchmark VLA performance against these results to assess the efficacy of synthetic robotization.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "Pretraining a VLA policy on HuRo data improves task completion from 51.5% to 80.3%."
          },
          {
            "claim_id": "1790095196747369430",
            "claim_ref": "5d4e99c84ea279b93eed333af3defba81de4912d",
            "snapshot": "{\"claim_id\":\"1790095196747369430\",\"claim_text\":\"Visual robotization improves OOD robustness and end-to-end pretraining with retargeted actions outperforms visual-only transfer.\",\"claim_type\":\"analysis\",\"confidence\":\"measured\",\"entities\":[{\"name\":\"HuRo\",\"role\":\"subject\",\"tag_id\":\"17891279667987836\",\"type\":\"topic\"}],\"evidence\":\"derived\",\"featured\":true,\"key_point_index\":4,\"quote\":null,\"signal\":\"Adopt retargeted action pretraining in VLA development to maximize OOD performance.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "Visual robotization improves OOD robustness and end-to-end pretraining with retargeted actions outperforms visual-only transfer."
          }
        ],
        "entities": [],
        "headline": "HuRo: Robotizing Human Videos for Scalable VLA Pretraining",
        "home_domain": null,
        "manifest_id": "1790093784543444754",
        "published_at": "2026-09-22",
        "significance": "high",
        "source_channel": "arXiv - Official Multimodal Document AI",
        "source_name": "arxiv-multimodal-document-ai",
        "source_url": "https://arxiv.org/pdf/2609.10706v4",
        "stream_id": null,
        "stream_ids": [
          "17779468859058016",
          "17894502471992576"
        ],
        "summary": "This paper introduces HuRo, a scalable pipeline for transforming human video data into robot-aligned observations and action trajectories. By leveraging 630K episodes of processed human video, the authors show that VLA policies can achieve substantial improvements in both task completion and out-of-distribution robustness. This approach provides a viable path for scaling robot learning without relying solely on expensive, task-specific real-robot data.",
        "tags": [
          "arXiv",
          "Multimodal AI"
        ]
      },
      {
        "brief": {
          "actionable_takeaways": [
            "Prioritize simple random masking strategies in OCL pipelines unless specific background segmentation improvements are required.",
            "Avoid complex attention-guided masking modifications for general object discovery tasks to prevent performance degradation.",
            "Use the provided code and evaluation logs to benchmark existing OCL architectures against these findings."
          ],
          "key_insights": [
            "AGM provides specific improvements in background segmentation on datasets with realistic textures (COCO and VOC) under unconditional slot initialization."
          ],
          "tldr": "Attention-guided masking provides no consistent performance benefit over random masking for object discovery in object-centric learning.",
          "unresolved": [
            "No clear performance advantage for AGM over RM across general object discovery tasks."
          ],
          "why_it_matters": "Informs researchers and AI engineers on effective self-supervision strategies for object-centric representation learning, preventing wasted compute on ineffective architectural modifications."
        },
        "claim_count": 3,
        "claims": [
          {
            "claim_id": "1790095204276513072",
            "claim_ref": "5a711111cfbbd60bba63a5e60482fa84f5ca3a80",
            "snapshot": "{\"claim_id\":\"1790095204276513072\",\"claim_text\":\"Attention-Guided Masking (AGM) does not consistently outperform Random Masking (RM) in object discovery accuracy.\",\"claim_type\":\"analysis\",\"confidence\":\"measured\",\"entities\":[{\"name\":\"multimodal_ai\",\"role\":\"subject\",\"tag_id\":\"17730948119041167\",\"type\":\"topic\"}],\"evidence\":\"observed\",\"featured\":true,\"key_point_index\":1,\"quote\":null,\"signal\":\"Researchers should monitor this finding to avoid adopting ineffective masking strategies in OCL model development.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "Attention-Guided Masking (AGM) does not consistently outperform Random Masking (RM) in object discovery accuracy."
          },
          {
            "claim_id": "1790095204283257830",
            "claim_ref": "e97dc290c18cd908c3d6158e3a08995a81848394",
            "snapshot": "{\"claim_id\":\"1790095204283257830\",\"claim_text\":\"AGM improves background segmentation on datasets with realistic textures (COCO and VOC) under unconditional slot initialization.\",\"claim_type\":\"data\",\"confidence\":\"measured\",\"entities\":[{\"name\":\"multimodal_ai\",\"role\":\"subject\",\"tag_id\":\"17730948119041167\",\"type\":\"topic\"}],\"evidence\":\"observed\",\"featured\":true,\"key_point_index\":2,\"quote\":null,\"signal\":\"Operators should consider AGM specifically for tasks requiring improved background segmentation in realistic texture environments.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "AGM improves background segmentation on datasets with realistic textures (COCO and VOC) under unconditional slot initialization."
          },
          {
            "claim_id": "1790095204287998468",
            "claim_ref": "f428ce9275ce92007b8d440bae769b659ae3f835",
            "snapshot": "{\"claim_id\":\"1790095204287998468\",\"claim_text\":\"Foreground object discovery remains comparable or decreases when using AGM across datasets.\",\"claim_type\":\"analysis\",\"confidence\":\"measured\",\"entities\":[{\"name\":\"multimodal_ai\",\"role\":\"subject\",\"tag_id\":\"17730948119041167\",\"type\":\"topic\"}],\"evidence\":\"observed\",\"featured\":true,\"key_point_index\":3,\"quote\":null,\"signal\":\"This serves as a performance warning for developers optimizing OCL models for foreground object detection.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "Foreground object discovery remains comparable or decreases when using AGM across datasets."
          }
        ],
        "entities": [],
        "headline": "Does Attention-Guided Masking Really Help Object Discovery in Object-Centric Learning?",
        "home_domain": null,
        "manifest_id": "1790093784543011548",
        "published_at": "2026-09-22",
        "significance": "medium",
        "source_channel": "arXiv - Official Multimodal Document AI",
        "source_name": "arxiv-multimodal-document-ai",
        "source_url": "https://arxiv.org/pdf/2609.15187v2",
        "stream_id": null,
        "stream_ids": [
          "17779468859058016",
          "17894502471992576"
        ],
        "summary": "The study investigates the efficacy of using attention-guided masking (AGM) to improve self-supervision in object-centric learning (OCL) models. Contrary to the intuition that attention maps provide better guidance for masked image reconstruction, empirical results across six datasets show that AGM fails to provide consistent performance gains over simple random masking (RM). The findings suggest that researchers should be cautious when attempting to leverage internal attention semantics for masked decoding, as it may lead to performance degradation in foreground object discovery.",
        "tags": [
          "arXiv",
          "Multimodal AI"
        ]
      },
      {
        "brief": {
          "actionable_takeaways": [
            "Implement validator-guided selective correction pipelines to optimize LLM/VLM inference costs in multimodal planning tasks."
          ],
          "key_insights": [
            "Perception-Evaluation-Revision chain improves L2, Collision, and Intersection metrics compared to the OmniDrive baseline.",
            "Validator-guided selective correction significantly reduces LLM/VLM token usage and generation costs.",
            "The framework provides a verifiable, auditable approach to autonomous driving supervision, enhancing safety and interpretability."
          ],
          "tldr": "A verifiable, cost-efficient framework for autonomous driving that uses a structured Perception-Evaluation-Revision chain to reduce reliance on expensive frontier model supervision.",
          "unresolved": [
            "Code, prompts, and validator scripts are not yet released."
          ],
          "why_it_matters": "Enables AI operators and developers to scale high-quality training data for autonomous systems while significantly reducing operational inference costs."
        },
        "claim_count": 3,
        "claims": [
          {
            "claim_id": "1790095247833802803",
            "claim_ref": "4c1be0d1a14376e563455726282a240794f60a34",
            "snapshot": "{\"claim_id\":\"1790095247833802803\",\"claim_text\":\"The Perception-Evaluation-Revision chain improves L2, Collision, and Intersection metrics over the OmniDrive baseline.\",\"claim_type\":\"data\",\"confidence\":\"measured\",\"entities\":[{\"name\":\"multimodal_ai\",\"role\":\"subject\",\"tag_id\":\"17730948119041167\",\"type\":\"topic\"}],\"evidence\":\"observed\",\"featured\":true,\"key_point_index\":null,\"quote\":null,\"signal\":\"Monitor this framework as a benchmark for planning-oriented vision-language models in autonomous driving.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "The Perception-Evaluation-Revision chain improves L2, Collision, and Intersection metrics over the OmniDrive baseline."
          },
          {
            "claim_id": "1790095247836815403",
            "claim_ref": "254b58b838864939aae94c6b14073a805c48b259",
            "snapshot": "{\"claim_id\":\"1790095247836815403\",\"claim_text\":\"Validator-guided selective correction reduces LLM/VLM token usage and generation time compared to full-sample generation.\",\"claim_type\":\"analysis\",\"confidence\":\"measured\",\"entities\":[{\"name\":\"multimodal_ai\",\"role\":\"subject\",\"tag_id\":\"17730948119041167\",\"type\":\"topic\"}],\"evidence\":\"observed\",\"featured\":true,\"key_point_index\":null,\"quote\":null,\"signal\":\"AI operators should evaluate this selective correction approach to optimize inference costs in high-volume data pipelines.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "Validator-guided selective correction reduces LLM/VLM token usage and generation time compared to full-sample generation."
          },
          {
            "claim_id": "1790095247838717878",
            "claim_ref": "aa73848ca08eb3eea499de7d58cd3405b1f294e9",
            "snapshot": "{\"claim_id\":\"1790095247838717878\",\"claim_text\":\"Structured revision targets improve vision-language planning supervision under realistic annotation budgets.\",\"claim_type\":\"statement\",\"confidence\":\"measured\",\"entities\":[{\"name\":\"autonomous_driving\",\"role\":\"subject\",\"tag_id\":\"17730926128324565\",\"type\":\"topic\"}],\"evidence\":\"observed\",\"featured\":true,\"key_point_index\":null,\"quote\":null,\"signal\":\"Researchers and builders should adopt structured revision targets to improve model performance while managing annotation costs.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "Structured revision targets improve vision-language planning supervision under realistic annotation budgets."
          }
        ],
        "entities": [],
        "headline": "VeriDrive: Verifiable Counterfactual Supervision for Cost-Efficient Vision-Language Planning",
        "home_domain": null,
        "manifest_id": "1790093784543906695",
        "published_at": "2026-09-22",
        "significance": "high",
        "source_channel": "arXiv - Official Multimodal Document AI",
        "source_name": "arxiv-multimodal-document-ai",
        "source_url": "https://arxiv.org/pdf/2606.07338v2",
        "stream_id": null,
        "stream_ids": [
          "17779468859058016",
          "17894502471992576"
        ],
        "summary": "VeriDrive addresses the high cost and free-form nature of reasoning supervision in vision-language driving models. It introduces a structured chain-of-thought process that grounds future motion, evaluates trajectories, and revises risky intents. This framework allows for more efficient data construction by escalating only difficult samples to frontier models, offering a scalable path for training planning-oriented agents.",
        "tags": [
          "arXiv",
          "Multimodal AI"
        ]
      },
      {
        "brief": {
          "actionable_takeaways": [
            "Use CytoCrowd to benchmark medical AI models against both noisy, real-world expert labels and clean gold standards.",
            "Incorporate annotation aggregation techniques to improve model robustness in domains where expert consensus is difficult to achieve."
          ],
          "facts": [
            {
              "label": "images",
              "value": "446"
            },
            {
              "label": "pathologists",
              "value": "4"
            }
          ],
          "key_insights": [
            "Introduces CytoCrowd, a dataset of 446 high-resolution cytology images with multi-pathologist annotations.",
            "Enables development and evaluation of annotation aggregation algorithms in medical image analysis."
          ],
          "tldr": "CytoCrowd provides a multi-annotator cytology dataset featuring both conflicting expert labels and a senior gold standard to improve medical AI robustness.",
          "why_it_matters": "Essential for developers and researchers building medical diagnostic AI that must account for subjective expert disagreement and high-stakes labeling variance."
        },
        "claim_count": 2,
        "claims": [
          {
            "claim_id": "1790095241707885991",
            "claim_ref": "38d236fc4c30a30276cd07aaec1cae2c4c023f46",
            "snapshot": "{\"claim_id\":\"1790095241707885991\",\"claim_text\":\"CytoCrowd provides 446 high-resolution cytology images with both raw, conflicting annotations from four pathologists and a senior expert gold standard.\",\"claim_type\":\"data\",\"confidence\":\"measured\",\"entities\":[{\"name\":\"CytoCrowd\",\"role\":\"subject\",\"type\":\"topic\"}],\"evidence\":\"direct_quote\",\"featured\":true,\"key_point_index\":null,\"quote\":\"The dataset features 446 high-resolution images, each with two key components: (1) raw, conflicting annotations from four independent pathologists, and (2) a separate, high-quality gold-standard ground truth established by a senior expert.\",\"signal\":\"Use for benchmarking medical AI models against noisy expert labels and evaluating annotation aggregation performance.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "CytoCrowd provides 446 high-resolution cytology images with both raw, conflicting annotations from four pathologists and a senior expert gold standard."
          },
          {
            "claim_id": "1790095241709923452",
            "claim_ref": "e28c4a004291290ad96ec68253a7bb19145ec87d",
            "snapshot": "{\"claim_id\":\"1790095241709923452\",\"claim_text\":\"The dual-structure of CytoCrowd allows for the evaluation of annotation aggregation algorithms in medical image analysis.\",\"claim_type\":\"statement\",\"confidence\":\"implied\",\"entities\":[{\"name\":\"CytoCrowd\",\"role\":\"subject\",\"type\":\"topic\"}],\"evidence\":\"derived\",\"featured\":true,\"key_point_index\":null,\"quote\":null,\"signal\":\"Monitor for new aggregation methods in medical imaging pipelines that leverage multi-annotator datasets.\",\"source_urls\":[],\"supporting_quotes\":[]}",
            "text": "The dual-structure of CytoCrowd allows for the evaluation of annotation aggregation algorithms in medical image analysis."
          }
        ],
        "entities": [],
        "headline": "CytoCrowd: A Multi-Annotator Benchmark Dataset for Cytology Image Analysis",
        "home_domain": null,
        "manifest_id": "1790093784543625410",
        "published_at": "2026-09-22",
        "significance": "medium",
        "source_channel": "arXiv - Official Multimodal Document AI",
        "source_name": "arxiv-multimodal-document-ai",
        "source_url": "https://arxiv.org/pdf/2602.06674v2",
        "stream_id": null,
        "stream_ids": [
          "17779468859058016",
          "17894502471992576"
        ],
        "summary": "CytoCrowd addresses the lack of ground-truth diversity in medical imaging by providing a dataset with both raw, conflicting expert annotations and a senior-verified gold standard. This dual-structure allows AI developers to benchmark standard object detection models while also training and evaluating algorithms designed to resolve expert disagreement. The dataset is a valuable resource for researchers and operators building medical diagnostic tools that require high reliability and robustness to subjective human labeling.",
        "tags": [
          "arXiv",
          "Zhao Chen",
          "Document AI"
        ]
      }
    ],
    "stories_per_domain": 24,
    "story_attempt": 1,
    "story_slot": 12,
    "story_units": [
      {
        "paragraphs": [
          {
            "citations": [
              {
                "claim_id": "1790095196719128958",
                "kind": "claim",
                "manifest_id": "1790093784543444754"
              }
            ],
            "text": "AI developers are scaling vision-language-action (VLA) policies by converting human video data into robot-aligned observations and action trajectories. This pipeline, called HuRo, bypasses the need to rely solely on expensive, task-specific real-robot data."
          },
          {
            "citations": [
              {
                "claim_id": "1790095196735321611",
                "kind": "claim",
                "manifest_id": "1790093784543444754"
              },
              {
                "claim_id": "1790095196739357477",
                "kind": "claim",
                "manifest_id": "1790093784543444754"
              },
              {
                "claim_id": "1790095196747369430",
                "kind": "claim",
                "manifest_id": "1790093784543444754"
              }
            ],
            "text": "The HuRo dataset comprises 630K robotized episodes and 142M processed frames. Pretraining a VLA policy on this data improves task completion from 51.5% to 80.3%, while visual robotization improves out-of-distribution robustness."
          }
        ],
        "single_source": false,
        "title": "Scaling Robot Learning with Human Video"
      },
      {
        "paragraphs": [
          {
            "citations": [
              {
                "claim_id": "1790097763129214790",
                "kind": "claim",
                "manifest_id": "1790097172664828006"
              }
            ],
            "text": "While general manipulation benefits from large-scale video pretraining, specialized medical procedures require high-fidelity, feel-based demonstration data. HumynexSurg-1 is the first curated expert liposuction dataset featuring synchronized force/torque, RGB-D video, and expert narration."
          },
          {
            "citations": [
              {
                "claim_id": "1790097763131447727",
                "kind": "claim",
                "manifest_id": "1790097172664828006"
              },
              {
                "claim_id": "1790097763133081034",
                "kind": "claim",
                "manifest_id": "1790097172664828006"
              }
            ],
            "text": "This specialized dataset allows rapid adaptation. NVIDIA Isaac GR00T N1.7 can be fine-tuned on the HumynexSurg-1 dataset in under one hour per run. Scaling analysis shows that every new session added to the training set lowers the error on unseen sessions."
          }
        ],
        "single_source": false,
        "title": "Specialized Multimodal Data for Surgical Robotics"
      }
    ],
    "summary": "Researchers are scaling robot learning through human video conversion and specialized datasets that enable rapid fine-tuning of foundation models for surgical tasks.",
    "supply": {
      "briefs": 268,
      "manifests": 311,
      "mcp_servable": 192,
      "signals": 268
    },
    "supporting_evidence": [
      {
        "home_domain": "engineering-technology",
        "manifest_id": "1790093784543303523",
        "relevance_reasons": [
          "same_source_host",
          "shared_stream",
          "shared_topics"
        ],
        "score": 6.3,
        "shared_tokens": [
          "arxiv",
          "multimodal"
        ],
        "stream_ids": [
          "17779468859058016",
          "17894502471992576"
        ]
      },
      {
        "home_domain": "engineering-technology",
        "manifest_id": "1790093784543444754",
        "relevance_reasons": [
          "same_source_host",
          "shared_stream",
          "shared_topics"
        ],
        "score": 6.3,
        "shared_tokens": [
          "arxiv",
          "multimodal"
        ],
        "stream_ids": [
          "17779468859058016",
          "17894502471992576"
        ]
      },
      {
        "home_domain": "engineering-technology",
        "manifest_id": "1790093784543011548",
        "relevance_reasons": [
          "same_source_host",
          "shared_stream",
          "shared_topics"
        ],
        "score": 6.252941,
        "shared_tokens": [
          "arxiv",
          "multimodal"
        ],
        "stream_ids": [
          "17779468859058016",
          "17894502471992576"
        ]
      },
      {
        "home_domain": "engineering-technology",
        "manifest_id": "1790093784543906695",
        "relevance_reasons": [
          "same_source_host",
          "shared_stream",
          "shared_topics"
        ],
        "score": 6.252941,
        "shared_tokens": [
          "arxiv",
          "multimodal"
        ],
        "stream_ids": [
          "17779468859058016",
          "17894502471992576"
        ]
      },
      {
        "home_domain": "engineering-technology",
        "manifest_id": "1790093784543625410",
        "relevance_reasons": [
          "same_source_host",
          "shared_stream"
        ],
        "score": 6.233333,
        "shared_tokens": [
          "arxiv",
          "dataset"
        ],
        "stream_ids": [
          "17779468859058016",
          "17894502471992576"
        ]
      }
    ],
    "supporting_home_domains": [
      "engineering-technology"
    ],
    "supporting_manifest_ids": [
      "1790093784543011548",
      "1790093784543303523",
      "1790093784543444754",
      "1790093784543625410",
      "1790093784543906695"
    ],
    "supporting_stream_ids": [
      "17779468859058016",
      "17894502471992576"
    ],
    "tail": [],
    "title": "New Datasets and Pipelines Accelerate Robotic Manipulation and Surgical Training",
    "window": {
      "hours": 12,
      "since": "2026-09-22T13:37:32.821807+00:00",
      "until": "2026-09-23T01:37:32.821807+00:00"
    }
  },
  "canonical_story_id": "a6e0808c43a6a023",
  "citation_ledger": {
    "assignment": {
      "assignment_desk_version": "v2",
      "assignment_rank": 12,
      "canonical_story_id": "hc_2fc7469e80ee3da77a3b3a81",
      "canonical_url": "/news/story/hc_2fc7469e80ee3da77a3b3a81/",
      "cluster_manifest_ids": [
        "1790097172664828006"
      ],
      "cluster_signature": "hcsig_v1_de7b8fbe6846de0394000de505b19fbc7b6ba713227fb8da9454728752508739",
      "cluster_stream_ids": [
        "17779468859058016",
        "17780442148971635",
        "17894412312039768"
      ],
      "cross_domain_evidence": false,
      "dedupe_reason": "global prewrite assignment",
      "evidence_manifest_ids": [
        "1790093784543011548",
        "1790093784543303523",
        "1790093784543444754",
        "1790093784543625410",
        "1790093784543906695",
        "1790097172664828006"
      ],
      "evidence_stream_ids": [
        "17779468859058016",
        "17780442148971635",
        "17894412312039768",
        "17894502471992576"
      ],
      "newsworthiness_score": {
        "audience_fit": 0.55,
        "breadth": 0.5125,
        "domain_priority": 0.98,
        "materiality": 0.9,
        "novelty": 1.0,
        "source_strength": 0.75,
        "total": 0.78725
      },
      "primary_home_domain": "engineering-technology",
      "secondary_home_domains": [
        "physical-sciences-mathematics"
      ],
      "supporting_evidence": [
        {
          "home_domain": "engineering-technology",
          "manifest_id": "1790093784543303523",
          "relevance_reasons": [
            "same_source_host",
            "shared_stream",
            "shared_topics"
          ],
          "score": 6.3,
          "shared_tokens": [
            "arxiv",
            "multimodal"
          ],
          "stream_ids": [
            "17779468859058016",
            "17894502471992576"
          ]
        },
        {
          "home_domain": "engineering-technology",
          "manifest_id": "1790093784543444754",
          "relevance_reasons": [
            "same_source_host",
            "shared_stream",
            "shared_topics"
          ],
          "score": 6.3,
          "shared_tokens": [
            "arxiv",
            "multimodal"
          ],
          "stream_ids": [
            "17779468859058016",
            "17894502471992576"
          ]
        },
        {
          "home_domain": "engineering-technology",
          "manifest_id": "1790093784543011548",
          "relevance_reasons": [
            "same_source_host",
            "shared_stream",
            "shared_topics"
          ],
          "score": 6.252941,
          "shared_tokens": [
            "arxiv",
            "multimodal"
          ],
          "stream_ids": [
            "17779468859058016",
            "17894502471992576"
          ]
        },
        {
          "home_domain": "engineering-technology",
          "manifest_id": "1790093784543906695",
          "relevance_reasons": [
            "same_source_host",
            "shared_stream",
            "shared_topics"
          ],
          "score": 6.252941,
          "shared_tokens": [
            "arxiv",
            "multimodal"
          ],
          "stream_ids": [
            "17779468859058016",
            "17894502471992576"
          ]
        },
        {
          "home_domain": "engineering-technology",
          "manifest_id": "1790093784543625410",
          "relevance_reasons": [
            "same_source_host",
            "shared_stream"
          ],
          "score": 6.233333,
          "shared_tokens": [
            "arxiv",
            "dataset"
          ],
          "stream_ids": [
            "17779468859058016",
            "17894502471992576"
          ]
        }
      ],
      "supporting_home_domains": [
        "engineering-technology"
      ],
      "supporting_manifest_ids": [
        "1790093784543011548",
        "1790093784543303523",
        "1790093784543444754",
        "1790093784543625410",
        "1790093784543906695"
      ],
      "supporting_stream_ids": [
        "17779468859058016",
        "17894502471992576"
      ]
    },
    "assignment_prefetch_billed_manifests": 61,
    "calls": [
      {
        "billed_manifests": 0,
        "called_at": "2026-09-23T01:37:56.259921+00:00",
        "elapsed_ms": 465.94,
        "manifest_ids": [],
        "mode": "count",
        "payload_bytes": 1013542,
        "tool": "synorb-manifests"
      },
      {
        "billed_manifests": 0,
        "called_at": "2026-09-23T01:37:56.495493+00:00",
        "elapsed_ms": 227.62,
        "manifest_ids": [],
        "mode": "count",
        "payload_bytes": 503508,
        "tool": "synorb-manifests"
      },
      {
        "billed_manifests": 40,
        "called_at": "2026-09-23T01:37:56.975223+00:00",
        "elapsed_ms": 460.97,
        "manifest_ids": [
          "1790097172664868273",
          "1790097172664839611",
          "1790097172664828006",
          "1790097172664222686",
          "1790097172664020194",
          "1790097172664018520",
          "1790097172663902782",
          "1790097172663683176",
          "1790097172663644988",
          "1790097172663634734",
          "1790097172663633390",
          "1790097172663575908",
          "1790097172663556723",
          "1790097172663435088",
          "1790097172663430532",
          "1790097172663428827",
          "1790097172663387216",
          "1790097172663280329",
          "1790097172663208903",
          "1790097172663169338",
          "1790097172663149144",
          "1790097172662830548",
          "1790097172662404413",
          "1790097172662283355",
          "1790097172662196656",
          "1790097172662141290",
          "1790093784543906695",
          "1790093784543834971",
          "1790093784543803320",
          "1790093784543625410",
          "1790093784543616238",
          "1790093784543602208",
          "1790093784543580142",
          "1790093784543538193",
          "1790093784543444754",
          "1790093784543353008",
          "1790093784543303523",
          "1790093784543011548",
          "1790093784542986388",
          "1790093784542958065"
        ],
        "mode": "default",
        "payload_bytes": 1495750,
        "tool": "synorb-manifests"
      },
      {
        "billed_manifests": 21,
        "called_at": "2026-09-23T01:37:57.839102+00:00",
        "elapsed_ms": 812.47,
        "manifest_ids": [
          "1790123338611015519",
          "1790118688414709549",
          "1790117309173031072",
          "1790115696962622042",
          "1790115215813444194",
          "1790115006079958854",
          "1790112657542883083",
          "1790110615160938026",
          "1790110183595050550",
          "1790107843615568372",
          "1790102009837394633",
          "1790100990136889009",
          "1790100990136512200",
          "1790100990136452882",
          "1790100990136340757",
          "1790100990136151262",
          "1790100990136064250",
          "1790099573510389318",
          "1790097172664868273",
          "1790097172664839611",
          "1790097172664828006",
          "1790097172664020194",
          "1790097172664018520",
          "1790097172663902782",
          "1790097172663683176",
          "1790097172663634734",
          "1790097172663575908",
          "1790097172663435088",
          "1790097172663428827",
          "1790097172663387216",
          "1790097172663280329",
          "1790097172663169338",
          "1790097172663149144",
          "1790097172662830548",
          "1790097172662404413",
          "1790097172662283355",
          "1790097172662141290",
          "1790095456839763188",
          "1790095456839547619",
          "1790095456838776941"
        ],
        "mode": "default",
        "payload_bytes": 1819905,
        "tool": "synorb-manifests"
      }
    ],
    "distinct_manifest_ids": [
      "1790093784543011548",
      "1790093784543303523",
      "1790093784543444754",
      "1790093784543625410",
      "1790093784543906695",
      "1790097172664828006"
    ],
    "edition_slot": "00",
    "excluded_manifest_ids": [],
    "prompt_version": "news_editorial_v2",
    "run_started_at": "2026-09-23T01:37:55.723687+00:00",
    "stories_per_domain": 24,
    "story_slot": 12,
    "strict_window": {
      "client_filtered": true,
      "published_date_from": "2026-09-22T13:37:32.821807+00:00",
      "published_date_to": "2026-09-23T01:37:32.821807+00:00"
    },
    "total_billed_manifests": 6,
    "total_calls": 4,
    "total_payload_bytes": 4832705
  },
  "cluster_manifest_ids": [
    "1790097172664828006",
    "1790093784543303523",
    "1790093784543444754",
    "1790093784543011548",
    "1790093784543906695",
    "1790093784543625410"
  ],
  "cluster_signature": "title:new datasets and pipelines accelerate robotic manipulation and surgical training",
  "corrections": [],
  "dedupe_reason": "canonical public story",
  "edition_date": "2026-09-23",
  "fact_claim_map": [
    {
      "claims": [
        {
          "claim_id": "1790095196719128958",
          "claim_ref": "72f40efb17977f613638929bd4c9017886681b56",
          "claim_text": "A robotization pipeline can convert heterogeneous human videos into robot-aligned observations and action trajectories.",
          "kind": "claim",
          "manifest_headline": "HuRo: Robotizing Human Videos for Scalable VLA Pretraining",
          "manifest_id": "1790093784543444754",
          "mcp_url": "https://synorb.com/agents?manifest_id=1790093784543444754&utm_source=hangingcontext&utm_medium=news_citation&utm_campaign=hc_news_public",
          "number": 1,
          "source_name": "arXiv - Official Multimodal Document AI",
          "source_url": "https://arxiv.org/pdf/2609.10706v4"
        }
      ],
      "paragraph": 1,
      "text": "AI developers are scaling vision-language-action (VLA) policies by converting human video data into robot-aligned observations and action trajectories. This pipeline, called HuRo, bypasses the need to rely solely on expensive, task-specific real-robot data."
    },
    {
      "claims": [
        {
          "claim_id": "1790095196735321611",
          "claim_ref": "87d152aef77a66a78d04c9e3f79a7df25146af96",
          "claim_text": "The HuRo dataset comprises 630K robotized episodes and 142M processed frames.",
          "kind": "claim",
          "manifest_headline": "HuRo: Robotizing Human Videos for Scalable VLA Pretraining",
          "manifest_id": "1790093784543444754",
          "mcp_url": "https://synorb.com/agents?manifest_id=1790093784543444754&utm_source=hangingcontext&utm_medium=news_citation&utm_campaign=hc_news_public",
          "number": 2,
          "source_name": "arXiv - Official Multimodal Document AI",
          "source_url": "https://arxiv.org/pdf/2609.10706v4"
        },
        {
          "claim_id": "1790095196739357477",
          "claim_ref": "015aa72a3568db05db64c126463d81c2659a1b31",
          "claim_text": "Pretraining a VLA policy on HuRo data improves task completion from 51.5% to 80.3%.",
          "kind": "claim",
          "manifest_headline": "HuRo: Robotizing Human Videos for Scalable VLA Pretraining",
          "manifest_id": "1790093784543444754",
          "mcp_url": "https://synorb.com/agents?manifest_id=1790093784543444754&utm_source=hangingcontext&utm_medium=news_citation&utm_campaign=hc_news_public",
          "number": 3,
          "source_name": "arXiv - Official Multimodal Document AI",
          "source_url": "https://arxiv.org/pdf/2609.10706v4"
        },
        {
          "claim_id": "1790095196747369430",
          "claim_ref": "5d4e99c84ea279b93eed333af3defba81de4912d",
          "claim_text": "Visual robotization improves OOD robustness and end-to-end pretraining with retargeted actions outperforms visual-only transfer.",
          "kind": "claim",
          "manifest_headline": "HuRo: Robotizing Human Videos for Scalable VLA Pretraining",
          "manifest_id": "1790093784543444754",
          "mcp_url": "https://synorb.com/agents?manifest_id=1790093784543444754&utm_source=hangingcontext&utm_medium=news_citation&utm_campaign=hc_news_public",
          "number": 4,
          "source_name": "arXiv - Official Multimodal Document AI",
          "source_url": "https://arxiv.org/pdf/2609.10706v4"
        }
      ],
      "paragraph": 2,
      "text": "The HuRo dataset comprises 630K robotized episodes and 142M processed frames. Pretraining a VLA policy on this data improves task completion from 51.5% to 80.3%, while visual robotization improves out-of-distribution robustness."
    },
    {
      "claims": [
        {
          "claim_id": "1790097763129214790",
          "claim_ref": "30575459e5c669ba4418cdbcd859d8f7b9edfe39",
          "claim_text": "HumynexSurg-1 is the first curated expert liposuction dataset featuring synchronized force/torque, RGB-D video, and expert narration.",
          "kind": "claim",
          "manifest_headline": "HumynexSurg-1: A Curated Expert Liposuction Dataset",
          "manifest_id": "1790097172664828006",
          "mcp_url": "https://synorb.com/agents?manifest_id=1790097172664828006&utm_source=hangingcontext&utm_medium=news_citation&utm_campaign=hc_news_public",
          "number": 5,
          "source_name": "arXiv Frontier AI Methods Select",
          "source_url": "https://arxiv.org/pdf/2609.23885v1"
        }
      ],
      "paragraph": 3,
      "text": "While general manipulation benefits from large-scale video pretraining, specialized medical procedures require high-fidelity, feel-based demonstration data. HumynexSurg-1 is the first curated expert liposuction dataset featuring synchronized force/torque, RGB-D video, and expert narration."
    },
    {
      "claims": [
        {
          "claim_id": "1790097763131447727",
          "claim_ref": "c564a6e56771d1f4bcc5a887d7e349b614ff7f02",
          "claim_text": "NVIDIA Isaac GR00T N1.7 can be fine-tuned on the HumynexSurg-1 dataset in under one hour per run.",
          "kind": "claim",
          "manifest_headline": "HumynexSurg-1: A Curated Expert Liposuction Dataset",
          "manifest_id": "1790097172664828006",
          "mcp_url": "https://synorb.com/agents?manifest_id=1790097172664828006&utm_source=hangingcontext&utm_medium=news_citation&utm_campaign=hc_news_public",
          "number": 6,
          "source_name": "arXiv Frontier AI Methods Select",
          "source_url": "https://arxiv.org/pdf/2609.23885v1"
        },
        {
          "claim_id": "1790097763133081034",
          "claim_ref": "7b96f807ce2bff8135ae3ea15edf535e49cebeef",
          "claim_text": "Every new session added to the training set lowers the error on unseen sessions.",
          "kind": "claim",
          "manifest_headline": "HumynexSurg-1: A Curated Expert Liposuction Dataset",
          "manifest_id": "1790097172664828006",
          "mcp_url": "https://synorb.com/agents?manifest_id=1790097172664828006&utm_source=hangingcontext&utm_medium=news_citation&utm_campaign=hc_news_public",
          "number": 7,
          "source_name": "arXiv Frontier AI Methods Select",
          "source_url": "https://arxiv.org/pdf/2609.23885v1"
        }
      ],
      "paragraph": 4,
      "text": "This specialized dataset allows rapid adaptation. NVIDIA Isaac GR00T N1.7 can be fine-tuned on the HumynexSurg-1 dataset in under one hour per run. Scaling analysis shows that every new session added to the training set lowers the error on unseen sessions."
    }
  ],
  "gate_report": {
    "checked_at": "2026-09-23T01:37:32.821807+00:00",
    "deterministic_pass": true,
    "findings": [],
    "initial_findings": [
      {
        "available_citable_manifests": 6,
        "cited_manifests": 1,
        "code": "thin_distinct_cited_manifests",
        "message": "stories with multiple citable manifests need cited evidence from at least two distinct manifests",
        "severity": "block"
      }
    ],
    "initial_status": "blocked",
    "judge": {
      "cost_usd": 0.002872,
      "model": "gemini-3.1-flash-lite",
      "sentences": [
        {
          "classification": "framing",
          "reason": "The sentence summarizes the provided claims regarding HuRo and HumynexSurg-1.",
          "sentence": "New Datasets and Pipelines Accelerate Robotic Manipulation and Surgical Training Researchers are scaling robot learning through human video conversion and specialized datasets that enable rapid fine-tuning of foundation models for surgical tasks.",
          "verdict": "entailed"
        },
        {
          "classification": "factual",
          "reason": "Supported by the claim that a robotization pipeline converts human videos into robot-aligned observations and action trajectories.",
          "sentence": "AI developers are scaling vision-language-action (VLA) policies by converting human video data into robot-aligned observations and action trajectories.",
          "verdict": "entailed"
        },
        {
          "classification": "factual",
          "reason": "Supported by the context of the HuRo pipeline's function in converting human video data.",
          "sentence": "This pipeline, called HuRo, bypasses the need to rely solely on expensive, task-specific real-robot data.",
          "verdict": "entailed"
        },
        {
          "classification": "factual",
          "reason": "Directly supported by the cited claim.",
          "sentence": "The HuRo dataset comprises 630K robotized episodes and 142M processed frames.",
          "verdict": "entailed"
        },
        {
          "classification": "factual",
          "reason": "Directly supported by the cited claims regarding task completion improvement and OOD robustness.",
          "sentence": "Pretraining a VLA policy on this data improves task completion from 51.5% to 80.3%, while visual robotization improves out-of-distribution robustness.",
          "verdict": "entailed"
        },
        {
          "classification": "framing",
          "reason": "The sentence provides context for the introduction of the HumynexSurg-1 dataset.",
          "sentence": "While general manipulation benefits from large-scale video pretraining, specialized medical procedures require high-fidelity, feel-based demonstration data.",
          "verdict": "entailed"
        },
        {
          "classification": "factual",
          "reason": "Directly supported by the cited claim.",
          "sentence": "HumynexSurg-1 is the first curated expert liposuction dataset featuring synchronized force/torque, RGB-D video, and expert narration.",
          "verdict": "entailed"
        },
        {
          "classification": "framing",
          "reason": "Supported by the claim that the dataset allows fine-tuning in under one hour.",
          "sentence": "This specialized dataset allows rapid adaptation.",
          "verdict": "entailed"
        },
        {
          "classification": "factual",
          "reason": "Directly supported by the cited claim.",
          "sentence": "NVIDIA Isaac GR00T N1.7 can be fine-tuned on the HumynexSurg-1 dataset in under one hour per run.",
          "verdict": "entailed"
        },
        {
          "classification": "factual",
          "reason": "Directly supported by the cited claim.",
          "sentence": "Scaling analysis shows that every new session added to the training set lowers the error on unseen sessions.",
          "verdict": "entailed"
        }
      ],
      "status": "pass",
      "tokens_in": 1089,
      "tokens_out": 788
    },
    "metrics": {
      "cited_manifests": 2,
      "claims_available": 18,
      "lead_checked": true,
      "paragraphs_checked": 4
    },
    "rewrite_attempted": true,
    "status": "passed",
    "version": "news_pr_d_v2"
  },
  "home_domain": "engineering-technology",
  "lastmod": "2026-09-23",
  "primary_home_domain": "engineering-technology",
  "schema_version": 1,
  "secondary_home_domains": [],
  "status": "draft",
  "summary": "Researchers are scaling robot learning through human video conversion and specialized datasets that enable rapid fine-tuning of foundation models for surgical tasks.",
  "suppressed_duplicate_of": null,
  "title": "New Datasets and Pipelines Accelerate Robotic Manipulation and Surgical Training",
  "url": "https://hangingcontext.com/news/engineering-technology/2026-09-23-00-12-robotic-multimodal-pretraining-datasets/"
}