LeRobot v0.7.0 Adds Language Commands, Dataset Streaming, Remote Inference and Four World Action Models; FSDP1 and W&B Are Replaced, the Old Async Module Removed in Favor of Remote/Policy Server, and Benchmarks Are Author-Reported

On November 5, 2026, the LeRobot team published "LeRobot v0.7.0: From Terabytes to Prompting Robots" on Hugging Face, with a GitHub release page that lists breaking changes. LeRobot is an open robotics library for data, training and running robot policies. The release rebuilds installation, training, inference and the hardware layer, so it is as much an upgrade warning as a feature list. All benchmark numbers below are reported by the project or by the upstream model authors, and we did not reproduce any of them.

The headline feature is language in the loop. With interactive mode turned on in the rollout tool, the robot connection and policy stay loaded between instructions, and the operator types commands while the robot moves: a command to change the subtask, a command to ask the policy what it sees, and a command to give a goal so the policy picks its own subtasks until the operator takes control back. The last two need a policy that generates text, such as FineART-VLA or Wall-OSS-0.5, or an external planner behind an OpenAI-compatible API. FineART, from Scale AI and Hugging Face, is a dataset of more than 40,000 bimanual episodes (1,718 hours) with more than 533,000 subtask annotations, 1.9 TB in LeRobot format.

The data path changed too. Streaming mode now fetches only the video byte ranges that training needs, from the Hub or a storage bucket, visits each frame once per epoch and resumes deterministically; the team reports 2.1 to 2.5 times more samples per second with four data-loader workers. LanceDB-backed datasets can be read in place with a global shuffle. Moving image augmentation to the GPU is reported to train ACT 2.8 times faster on an H100 and 3.3 times faster on an A100. The release notes add a quiet but important change: several augmentations, including motion blur, random shadow and gamma correction, were silent no-ops on uint8 frames before and now actually apply, so retrained policies may behave differently for that reason alone.

Training was rebuilt for world action models of 5 to 20 billion parameters. FSDP1 is replaced by FSDP2 and HSDP sharding with distributed checkpoints. On an 8-GPU H200 node with a 6-billion-parameter FastWAM, the project reports checkpoint saves three times faster and a resume peak that falls from 143 GB to 42 GB per GPU. Four world action models and four vision-language-action models are added. The reported LIBERO results are close to the ceiling: Cosmos 3 Nano (16.5 billion parameters) at 97.90% and Cosmos 3 Edge (4.6 billion) at 97.95%, LaWAM at 98.5% (1,970 of 2,000 episodes). Long-WAM reached 99.00% in IDM and 98.45% in COD, each over 2,000 episodes. LingBot-VLA 2.0 reports 93.5% on RoboTwin 2.0 clean against 82.7% for pi-0.5 upstream, and with CUDA graphs and torch.compile an action chunk takes about 120 ms instead of about 756 ms on an RTX PRO 6000 in bf16. The G0.5 model is released under a non-commercial community license, not Apache 2.0.

Remote inference lets a laptop or single-board computer sit beside the robot while a GPU workstation runs the policy. Cameras, motors and recording stay local, observations go out, and chunks of actions come back; the client aligns and blends overlapping chunks, which hides inference latency. Evaluation reporting also improves: the evaluation tool now prints success counts with a 95% Wilson confidence interval, and the post gives the example of 9 of 10 episodes as 90.0% with an interval from 59.6% to 98.2%. That interval is a useful reminder that a few tenths of a point on a nearly saturated benchmark carry little information.

The breaking changes matter more to existing users than any single feature. Weights & Biases is replaced by Trackio, and the command-line options change from wandb to trackio. FSDP1 is gone, and FSDP checkpoints saved with 0.6.x and earlier cannot be resumed, although their weights still load. The asynchronous inference module and its install extra are removed in favor of the remote extra. Passing the PEFT option alone now trains LoRA, where it used to fully fine-tune silently. On Linux, torch and torchvision now come from PyPI as CUDA 13 builds that need NVIDIA driver 580.65 or newer. Simulator modules moved, several environment options were removed, and the Feetech and Dynamixel motor buses were rewritten on a new serial library, with old classes kept only as deprecated aliases. The post adds new hardware support for the I2RT YAM arm, the reBot B601 and a 23-degree-of-freedom Unitree G1, and says the Hub now holds almost 300 TB of open Physical AI data in this format.

Our reading: the release is a large, fast-moving upgrade, and the sensible approach is to read the breaking-changes list before bumping a dependency. Teams with FSDP checkpoints, Weights & Biases dashboards or asynchronous inference scripts have concrete migration work. The model results are close to the benchmark ceiling, were produced by the model authors, and say little about performance on a new robot, so judge the models on your own tasks. The interactive and remote features are the part most worth trying early, because they change how a robot is operated day to day.