Robot Brains Are Getting Better Data, Better Views, and a Sense of Done

The data engine

In a preprint, Flora is an open-source industrial data engine for robotics foundation models, according to the authors. It provides captioning, outcome reward labeling, process reward labeling, and pretraining data curation, and it is used for training GR00T models. The authors open-sourced the data engine to enable scaling of custom robotics datasets. m:…4818

The scale figures reported are large. Flora labels 200,000 hours of data per day, according to the authors, and utilizes 512 GB300s for its operations. Flora is roughly 125 times cheaper than GPT-6 Astra, according to the authors, and is competitive with GPT-6 Astra on internal benchmarks — a comparison the authors themselves qualify as internal, which is the right caution for any cost claim. m:…4818

Flora achieves state-of-the-art scores on the WGO-Bench, RoboReward, and Robometer benchmarks, according to the authors. For teams building robot datasets, the relevant claim is not any single benchmark but the combination: captioning, reward labeling, and curation in one engine, at a throughput that makes large-scale robot data practical to produce rather than only to buy. m:…4818

Looking where the action will be

In a second preprint, FocusVLA addresses a narrower problem in vision-language-action models: which camera view to use, and when. Existing VLA models tend to persistently rely on the wrist view, according to the authors, which limits stage adaptability and robustness across the phases of a manipulation task. m:…7317

FocusVLA predicts future interaction intent from historical actions to guide dynamic view routing, according to the authors, shifting between global and local views as the task stage changes. Its intent prediction and routing modules can be integrated into existing VLA frameworks without modifying their core architectures, according to the authors, which lowers the adoption cost for teams that already have a working VLA stack. m:…7317

In real-world evaluation, FocusVLA improved the success rate from 71.7% to 85.0%, according to the authors, and it improved the average success rate on RoboTwin 2.0 by 10.10 percentage points over the pi-0.5 baseline. The gap between benchmark and real-world figures is itself informative: view routing helps in both settings, but the real-world gain is the one that matters for deployment. m:…7317

Knowing when to stop

In a third preprint, Echo gives VLA models a sense of task status. Echo is a framework for Vision Language Action models that integrates stop and failure signals into VLA models to guide robot execution and recovery, according to the authors. It represents task status with stop and failure signals added as two dimensions of each action chunk. m:…2950

Existing methods for task status prediction often require fine-grained supervision, external monitoring models, or complex architectural designs, according to the authors. Echo instead constructs successful, failed, and recovery trajectories with complementary supervision properties to learn signals without fine-grained progress annotations, and it denoises status signals faster than robot actions so that action denoising can condition on cleaner status estimates. m:…2950

Real-world manipulation experiments show that Echo accurately predicts task status and consistently improves policy performance, according to the authors. A robot that knows it has failed can retry or ask for help; a robot that cannot tell success from failure will keep executing a broken plan. Status, in this framing, is not a dashboard metric but part of the action itself. m:…2950

What deployment still needs

The three papers share a premise: robot capability now depends as much on data pipelines, view selection, and status awareness as on the underlying model. FLORA was open-sourced to enable scaling of custom robotics datasets, according to the authors, and FocusVLA's routing modules integrate into existing frameworks without core changes — both choices lower the cost of trying these ideas on a team's own robots. m:…4818 m:…7317