LeRobot 0.7.0 adds world action models, remote inference and a rebuilt training stack
- Source: Hugging Face (LeRobot team)
- Author: Open Source Robotics
- Published: 2026-11-05
- Category: research
Hugging Face released LeRobot v0.7.0 on 5 November. It puts interactive rollouts in a release, adds four world action models and four vision-language-action models, rebuilds training for multi-billion-parameter models, and lets a robot run its policy on a remote GPU. The upgrade from 0.6 changes installation, experiment tracking and the motor drivers.
LeRobot, Hugging Face's open-source robot-learning library, released v0.7.0 on 5 November, three months after v0.6.1. The team's release post describes it as working on both ends of a policy's life: training on terabytes of demonstrations without downloading them first, and talking to the robot once the policy is running. It also rebuilds installation, training, inference and the hardware layer, so an upgrade from 0.6 needs some changes (see the last section).
Talking to a running policy
Interactive rollouts, which we covered when they reached the main branch, are now in a release. Add --interactive=true to lerobot-rollout and the robot, policy and processors stay loaded: /subtask changes the instruction while the arm moves, /vqa asks the policy what it sees, and /autosteer gives it a goal so it picks its own subtasks. The last two need a policy that writes text, such as FineART-VLA, trained on the FineART dataset, or Wall-OSS-0.5, or an external vision-language model behind an OpenAI-compatible API.
New models
Four world action models, which predict future video and the next actions together, join the library:
- Cosmos 3 from NVIDIA, in a 16.5B Nano and a 4.6B Edge size. The team reports 97.90% and 97.95% average success on LIBERO, and either can be fine-tuned in full or with LoRA. - FLUX 3 Action from Black Forest Labs. LeRobot runs inference and LoRA fine-tuning; training a new embodiment uses Black Forest Labs' own flux-action repository. - LaWAM from the RLinf team, which learns dynamics in a latent space. - Long-WAM from NVIDIA's LongLive team, built for long context.
Four vision-language-action models arrive too: OpenGalaxea's G0.5, which writes out its current subtask before its actions; Robbyant's LingBot-VLA 2.0; Dexmal's DM0.5, ported from OpenDM; and X Square Robot's Wall-OSS-0.5.
Training and data at scale
- FSDP1 is replaced by FSDP2 and HSDP sharding, gradient accumulation and distributed checkpoints. On an 8×H200 node with a 6B model, the team measured checkpoints saving 3x faster and the memory peak on resume falling from 143 GB to 42 GB per GPU (multi-GPU docs). - Training gains FP16 mixed precision, an optional moving average of the weights (EMA), and Trackio, Hugging Face's open-source tracker, which replaces Weights & Biases entirely. - Image augmentation can run on the GPU; the team measured ACT training 2.8x faster on an H100 and 3.3x on an A100. - Training from the Hub without a download (our story) and from LanceDB tables (our story) are both in the release.
Remote inference, simulation and hardware
- lerobot-policy-server runs the model on a GPU machine while a laptop or single-board computer beside the robot keeps the cameras, motors and recording. The robot follows buffered actions while the server computes the next chunk, over Zenoh (docs). - Every simulator now runs through one core, in-process or in its own Docker image behind lerobot-env-server, so simulators no longer fight over dependency versions. The first benchmark on it is NVIDIA's RoboLab, 120 tabletop tasks for the DROID Franka setup on Isaac Lab. lerobot-eval now reports a 95% confidence interval with every success rate. - Hardware: one or two I2RT YAM arms; the reBot B601 on RobStride as well as Damiao motors; and the Unitree G1 running the decoder of NVIDIA's SONIC whole-body controller, which we covered earlier, plus the 23-DoF G1.
Before you upgrade
- On Linux, torch now comes from PyPI as a CUDA 13 build that needs NVIDIA driver 580.65 or newer; with an older driver, install torch from a CUDA 12 index first (installation guide). macOS needs version 14 or later. - --wandb. options become --trackio., and checkpoints are no longer uploaded as artifacts by default. - FSDP checkpoints saved with 0.6 or earlier cannot be resumed, though their weights still load. - lerobot[async] and lerobot.async_inference are gone; install lerobot[remote] instead. - Feetech and Dynamixel motors now share one rustypot-based bus. The old class names still work as deprecated aliases, and calibration files are unchanged. - --policy.use_peft=true now trains LoRA; before, it silently ran a full fine-tune.
Why it matters
LeRobot is where many open robot policies, datasets and arms meet, so what it supports decides what most hobbyist and lab setups can try without writing their own glue. This release brings the newest model families, world action models among them, into the same training and rollout commands as older policies, and remote inference means a small robot computer no longer limits which model it can run. The breaking changes matter just as much: anyone with a working 0.6 setup should read the upgrade notes before updating.
Key takeaways
- LeRobot v0.7.0 came out on 5 November, three months after v0.6.1.
- Interactive rollouts are now in a release: change a running policy's instruction, ask what it sees, or give it a goal.
- Four world action models (Cosmos 3, FLUX 3 Action, LaWAM, Long-WAM) and four VLAs (G0.5, LingBot-VLA 2.0, DM0.5, Wall-OSS-0.5) are supported.
- lerobot-policy-server runs the model on a GPU machine while the robot's own computer handles cameras and motors.
- Upgrading changes installation (CUDA 13 builds), replaces Weights & Biases with Trackio, and moves Feetech and Dynamixel motors to one rustypot bus.
Open-source angle
LeRobot itself is Apache-2.0, but the models it now loads come under different terms: LingBot-VLA 2.0 is Apache-2.0 and the Cosmos 3 checkpoints use the OpenMDW 1.1 licence, while G0.5 is non-commercial, DM0.5 and FineART-VLA carry Gemma's terms, and FLUX 3 Action needs an access request under Black Forest Labs' own FLUX Kommunity License. Check the model card before building on a checkpoint.
Related technologies
- World action models
- Vision-language-action models
- FSDP2
- Remote inference
- Zenoh
Related projects
Links
- Source: https://huggingface.co/blog/lerobot/lerobot-release-v070