LeRobot v0.7.0 adds remote inference, Hub streaming and language-steered rollouts

Hugging Face released LeRobot v0.7.0 on 5 November 2026. The largest change for people with hardware on the bench is remote inference. A new lerobot-policy-server runs the policy on a GPU machine while a laptop or single-board computer next to the robot handles cameras, motors and recording. The client keeps executing buffered actions while the next chunk is computed, then blends the overlap. Transport uses Zenoh. This replaces the old async_inference module, which is removed.

Training can now stream datasets from the Hub or a storage bucket instead of downloading them, and datasets can be stored as LanceDB tables. Hugging Face reports that GPU image augmentation makes ACT train 2.8 times faster on an H100. The release adds four world action models and four vision-language-action (VLA) policies. One of them, G0.5, ships under a non-commercial licence rather than Apache-2.0. An --interactive rollout mode lets you change the subtask or query the policy while the robot runs. lerobot-eval now reports 95% Wilson confidence intervals alongside success counts.

The list of breaking changes is long. On Linux, PyTorch now comes as CUDA 13 builds that need NVIDIA driver 580.65 or newer. Weights & Biases is replaced by Trackio. FSDP checkpoints from v0.6 cannot be resumed. Feetech and Dynamixel motors move to a single SerialMotorsBus backed by rustypot instead of the vendor SDKs.

Why it matters

With remote inference, the computer beside a low-cost arm such as the SO-101 no longer needs the memory to load a large policy. The confidence intervals in lerobot-eval matter as much: the release post’s own example of 9 successes in 10 episodes spans 59.6% to 98.2%. Before upgrading a working rig, check the NVIDIA driver version, re-test motor calibration under the new bus code, and read the release notes in full. Keep v0.6 installed in a separate environment until your own recordings replay correctly.