Robots need data, ears, and a doorstep before they need bigger models

The data problem, labeled step by step

Scale AI and Hugging Face released FineART, which Scale describes as the largest open robot dataset with subtask labels. The dataset contains 40,543 recorded episodes, representing 1,718 hours of footage across 151 different tasks, and 533,913 labeled steps, with episodes labeled with both task-level and temporally segmented subtask annotations. The subtask labels are the point. A robot that has only ever seen a whole task succeed or fail cannot tell which step went wrong; a dataset that labels the steps gives a model something to learn from when it fails halfway through. m:…4028

The company says 77% of the data covers pick-and-place and sorting tasks, while 23% covers complex manipulation tasks including textiles, cables, tool use, and precision insertion, and that collection and validation involved a three-stage pipeline of collection, subtask annotation, and quality validation. The split is honest about where the field stands: most of what robots can be trained on at scale is still moving objects from one place to another. FineART-VLA, described as the first open System 1-System 2 model in LeRobot that can reason over step-by-step actions, uses System 2 planning to reason over scenes and predict subtasks and System 1 control to translate predicted subtasks into actions, supporting spatial instruction following, long-horizon execution, and cross-embodiment transfer. Both the dataset and the model are natively integrated into the Hugging Face LeRobot ecosystem, which means other teams can try to reproduce the results rather than take them on trust. m:…4028

Hearing where a sound comes from

Edge Sound Research, the company says, has raised $2M in new funding, bringing its total to $9M, to finish integrations with enterprise customers. The company builds audio infrastructure for physical AI systems, and its technology isolates sounds in noisy environments and pairs them with sensor data. The problem it names is easy to overlook: microphones record by location, which makes it difficult for AI to identify sound sources. A robot with ears that cannot tell where a sound came from is guessing at half its environment. m:…3807

The company's answer, it says, is a platform that uses digital twins of sensors to create a spatial map of the environment, built as software and models that are hardware-independent. Its proving ground so far is sports: the company says its Sound Engine is deployed across the sports industry, counts Shure, the National Basketball Association, Fox Sports, and the United States Tennis Association as customers and partners, and that sports broadcasting requires audio processing results in under 33 milliseconds. A latency budget that tight is a useful filter. If the system works under broadcast conditions, it has a plausible claim to work on a factory floor; if it only works in a quiet lab, it does not. m:…3807

A place to hand things over

Arrive AI's announcement is smaller in scale and, in its own way, more revealing. The company says it installed two Arrive Point AP3+ units in a commercial office building in Central Indiana, its first Arrive Point installation in a commercial office building, situated near the employee entrance for access by delivery personnel and autonomous mobile robots. The units provide a secure place for deliveries to be held until collection, and the company says it is installing them to learn how office building occupants use secure delivery points. This is infrastructure as an experiment: deploy the endpoint first, watch how people actually use it, then decide what to build next. m:…5270

The company says its platform supports drones, ground robotics, and human couriers, and that the AP3+ is designed for use on the Arrive Point Network. Read together with the other two stories, the pattern is clear. Physical AI does not yet fail mainly for lack of a cleverer model. It fails for lack of labeled examples of doing the job, lack of hearing that locates sounds in space, and lack of mundane endpoints where a machine can hand something to a person. The teams fixing those three gaps are building the layer the models will run on, and that layer is where the next advantage will sit. m:…5270 m:…4028 m:…3807