Hugging Face LeRobot v0.7.0 lets you talk to a running robot and adds world action models

The headline feature is interactive rollouts. Add a flag to lerobot-rollout and the robot connection and policy stay loaded between instructions, so you can drive the session from a terminal while the arm keeps moving[1]. /subtask swaps the current instruction, /vqa asks the policy what the robot can see, and /autosteer hands it a goal and lets it choose its own subtasks until you take back control.

Asking questions and autosteering need a policy that can generate text, such as FineART-VLA or Wall-OSS-0.5. If yours can't, LeRobot can call any vision language model behind an OpenAI-compatible API as a planner, turning a vague request like "tidy up" into precise steps the policy understands[1]. FineART itself, built by Scale AI and Hugging Face, is a dataset of more than 40,000 bimanual episodes (1,718 hours) with over 533,000 subtask labels.

v0.7.0 adds four world action models, which predict future video and the next actions together[1]:

• Cosmos 3 from NVIDIA, in Nano (16.5B) and Edge (4.6B) sizes, reporting about 98% average success on the LIBERO benchmark.
• FLUX 3 Action from Black Forest Labs, the FLUX image team's first move into robotics, with base, SO-101 and DROID checkpoints.
• LaWAM from RLinf, which learns dynamics in latent space and scores 98.5% on its LIBERO fine-tune.
• Long-WAM from NVIDIA's LongLive team, built for long context and landing around 99% on LIBERO.

Four vision-language-action models join too: OpenGalaxea's G0.5, Robbyant's LingBot-VLA 2.0, Dexmal's DM0.5 and X Square Robot's Wall-OSS-0.5[1]. Hugging Face flags that G0.5 uses a non-commercial licence, unlike the Apache-2.0 library around it[1].

The Hub now holds almost 300 TB of open robot data in LeRobot format, much of it added in recent months, from humanoid teleoperation in real homes to 100,000 hours of egocentric human video[1]. LeRobot can now stream those datasets straight from the Hub, a Storage Bucket or LanceDB tables instead of copying them to disk first, and image augmentation can move to the GPU, which Hugging Face says trains its ACT policy 2.8 times faster on an H100.

To fit 5 to 20 billion parameter models, the old FSDP1 path is gone in favour of FSDP2, HSDP and Torch distributed checkpoints. On an 8 by H200 node with a 6B model, checkpoint saves are three times faster and peak memory on resume drops from 143 GB to 42 GB per GPU[1].

A new lerobot-policy-server lets a laptop or single-board computer sit beside the robot while the model runs on a GPU workstation across the lab. The robot keeps following buffered actions while the server computes the next chunk, which hides inference latency[1]. Every simulator now runs through one shared core, locally or in its own Docker image, and lerobot-eval reports 95% confidence intervals so ten lucky episodes no longer look like a solved benchmark.

On the hardware side, the release adds the I2RT YAM arm, Seeed Studio's reBot B601, NVIDIA's SONIC whole-body controller on the Unitree G1, and the 23-DoF G1[1].

Open robotics has had plenty of models but few shared tools for running them. LeRobot is turning into that common layer: the place where NVIDIA, Black Forest Labs and a long list of labs drop their robot models and datasets so anyone can fine-tune them on cheap hardware. Prompting a robot mid-task, and running a 16B model on a GPU server instead of the robot itself, moves home and lab robotics a step closer to something hobbyists and startups can actually use.