THE ESSENTIALS IN 30 SECONDS

  • LeRobot v0.7.0, released on November 5, adds an interactive rollout mode where operators send text commands such as subtasks or questions to a running policy.
  • Datasets can now be streamed from the Hub without local download, and a new policy server runs models on a remote GPU while the robot stays local.
  • Training moves to FSDP2 and Trackio, and the Hub now holds almost 300 TB of open robot data in LeRobot format, according to Hugging Face.

Hugging Face published version 0.7.0 of LeRobot, its open-source library for robot learning, on November 5. The release centres on giving operators language-level control over policies that are already running on a robot, while also reworking how datasets are loaded and how large models are trained and served.

Talking to a running policy

A new interactive mode for the lerobot-rollout command keeps the robot connection and model weights loaded between instructions. Operators can type commands to assign a subtask, ask the policy a visual question, or hand control to an "autosteer" loop, and sessions can be recorded as new episodes. An external vision-language model, served through any OpenAI-compatible API, can act as the planner that answers questions and issues those steering prompts.

To support this kind of training, the team released FineART, a dataset built with Scale AI containing more than 40,000 bimanual episodes, 1,718 hours of recordings and over 533,000 subtask annotations. A companion model, FineART-VLA, extends the π0.5 policy with a language head. Annotation recipes in YAML turn labels into training targets.

Data, training and inference changes

Datasets can now be streamed from the Hugging Face Hub by fetching only the video byte ranges a training run needs; the blog reports 2.1 to 2.5 times more samples per second with four data-loading workers. A LanceDB backend and GPU-side image augmentation were also added.

On the training side, the older FSDP1 path has been replaced by FSDP2, hybrid sharding and distributed checkpoints. Hugging Face says that on an eight-GPU H200 node, peak memory when resuming a 6-billion-parameter model fell from 143 GB to 42 GB per GPU. Experiment logging switches from Weights & Biases to Trackio, which means existing command-line flags change. A new lerobot-policy-server runs the model on a remote GPU machine while cameras, motors and recording stay on the robot; it replaces the previous asynchronous inference module, which was removed.

Hardware, models and caveats

New hardware support includes I2RT's YAM arm in single or bimanual setups, a Seeed Studio arm and further Unitree G1 humanoid integrations. The release also adds several third-party world-action and vision-language-action models, including NVIDIA Cosmos 3 checkpoints. According to the blog, the Hub now hosts almost 300 TB of open physical-AI data in LeRobot format.

Most performance figures come from Hugging Face's own tests. Some newly integrated models carry their own licences, including at least one non-commercial licence, and Linux installs now require recent NVIDIA drivers for the CUDA 13 builds.