Breaking02 / IT
LeRobot v0.7.0 changes install, training and inference setup

Hugging Face released LeRobot v0.7.0, a new version of its robot-learning library, on GitHub on 5 Nov 2026. According to the release notes, the install, training and inference layers have changed, so users of existing options must revise their settings.
The LeRobot team said it added interactive sessions. These let users question a running robot or change its subtask. Cosmos 3, Long-WAM, FLUX 3 Action and LaWAM were added as world action models. Such models predict the future to decide on actions. Training can run directly from the Hub, a Storage Bucket or LanceDB, without downloading the dataset first, the team said.
Compatibility. According to the release notes, torch and torchvision on Linux now come from PyPI, the Python package index, as CUDA 13 builds. Those builds require NVIDIA driver 580.65 or newer. Trackio replaces the experiment tracker Weights & Biases, so --wandb.* is now --trackio.*.
The lerobot[async] extra is removed and replaced by lerobot[remote]. Multi-GPU parallel training has moved from FSDP1 to FSDP2 and HSDP, the release notes say. FSDP checkpoints saved with lerobot 0.6.x or earlier cannot be resumed.
G0.5, one of the new models, is not distributed under Apache-2.0, according to the release post. Its code and weights are covered by the non-commercial G0.5 Community License.
Where it runs. With remote inference, the model runs on a GPU workstation and a laptop or single-board computer can sit beside the robot, the LeRobot team said.
The release post does not give commercial terms for the new models other than G0.5.