NVIDIA releases Cosmos 3 Edge for on-device robot control

NVIDIA has introduced Cosmos 3 Edge, a 4B parameter model designed for on-device robot control. Part of the Cosmos 3 family, the model is optimized to run on NVIDIA Jetson Thor hardware, addressing constraints such as device memory and control latency. The model utilizes pre-trained physical-world data to predict robot actions, allowing for manipulation policies that can be evaluated in closed-loop simulations.

In parallel, new research highlights the complexities of robot learning data quality. The development of RDA (Robot Data Audit), an open-source tool, aims to address the gap between data structural integrity and behavioral usability. An audit of 4,959 episodes across 12 local datasets revealed that datasets passing basic loading and schema checks may still contain excessive idle motion, action discontinuities, or sampling problems that hinder effective robot training.