NVIDIA and Hugging Face expand LeRobot with Cosmos 3 and 'Data for Agents' initiative

NVIDIA and Hugging Face deepened their open-robotics collaboration, releasing LeRobot v0.6.0 with NVIDIA contributing Cosmos 3 vision models for embodied AI, and launching a joint 'Data for Agents' initiative to provide open datasets for AI-agent development. The partnership also brings NVIDIA's Isaac GR00T 1.7 reasoning vision-language-action (VLA) model and the Isaac Teleop framework into Hugging Face's open-source LeRobot library.

Mechanically, the goal is to standardize and democratize end-to-end robot development: LeRobot provides the open training/deployment framework, NVIDIA supplies the foundation VLA models and simulation/teleoperation tooling, and the shared data initiative addresses the field's biggest bottleneck — the scarcity of robotics training data. Hugging Face also demonstrated low-latency speech-to-speech pipelines powering 9,000+ Reachy Mini robots in partnership with Cerebras.

Separately, Hugging Face launched ML Intern, an open-source agentic model scoring 32% on GPQA versus Claude Code's 23% — a benchmark result HN robotics threads flagged as real progress for open agentic models, while questioning embodied-AI deployment timelines.

Competitively, this consolidates the open-source robotics stack around the NVIDIA-Hugging Face axis, positioning it against closed robotics efforts and reinforcing the week's physical-AI theme alongside Mistral's Robostral Navigate. Clement Delangue separately highlighted that Hugging Face Transformers models now run in vLLM at native speed, another open-inference unlock. Skeptics note benchmark wins don't guarantee real-world robot reliability. Watch adoption of Cosmos 3 and GR00T 1.7 by robotics teams and whether the open data initiative attracts meaningful dataset contributions.