LeRobot v0.7.0 turns the open robotics stack from Hugging Face into something you can talk to, train on without disk limits, and run across two machines. The release landed on GitHub on November 5, 2026, and the team's release blog is titled "From Terabytes to Prompting Robots." The GitHub release notes also list a long set of breaking changes, so read them before you upgrade a working setup.
This post covers what is new, what the numbers are and who reported them, what breaks, and how to try it.
TL;DR: LeRobot v0.7.0 at a glance #
| Question | Answer |
|---|---|
| When was it released? | November 5, 2026 (GitHub release v0.7.0, marked Latest) |
| Who made it? | Hugging Face and 77 contributors listed on the GitHub release page |
| Biggest new feature? | Interactive rollouts: steer a running robot with text commands |
| Training change? | Disk-less streaming from the Hub, a Storage Bucket or LanceDB |
| Inference change? | lerobot-policy-server runs the model on a remote GPU machine |
| New models? | Cosmos 3, FLUX 3 Action, LaWAM, Long-WAM, G0.5, LingBot-VLA 2.0, DM0.5, Wall-OSS-0.5, FineART-VLA |
| New hardware? | YAM arms, reBot B601, Unitree G1 with NVIDIA SONIC decoder |
| Will it break my setup? | Likely yes: CUDA 13, Trackio, removed async module, moved simulator imports |
| Is it free? | The library is open source. Individual models carry their own licenses |
What are interactive rollouts? #
Interactive rollouts are the headline feature. Run lerobot-rollout --interactive=true and the robot, the weights and the processors stay loaded between commands. You then steer the robot with a chat-style command line instead of restarting a script for each task.
The release blog lists three commands:
/subtaskchanges the instruction in the middle of a run./vqaasks the robot what it sees./autosteergives a goal that the policy breaks into its own subtasks.
The last two need a policy that can generate text, such as FineART-VLA or Wall-OSS-0.5, or an external vision-language model planner behind an OpenAI-compatible API. A plain action policy cannot answer questions. The GitHub changelog also lists "interactive sessions" for rollout control from code, so the same loop works programmatically.
Why it matters: in older versions, changing a task meant re a large model. For a 4B or larger policy that costs real time on every try. Keeping state loaded makes tuning prompts on real hardware much faster. It also makes the robot loop look like the agent loops that readers already know from coding tools. For background on that pattern, see our guide to loop engineering.
How does disk-less dataset streaming work? #
Robot datasets are large. Hugging Face says the Hub now holds almost 300 TB of open Physical AI data in LeRobot format. Version 0.7.0 lets you train from the Hub, a Storage Bucket or LanceDB without down the full set first.
Switch it on with --dataset.streaming=true. The GitHub notes say this uses a new disk-less episode-pool reader with its own --dataset.streaming_* options, and the old Backtrackable class is gone. You can also load a LanceDB set directly: LeRobotDataset("lance-format/pusht-lance").
Reported numbers, all from the Hugging Face release blog, not an independent test:
| Change | Reported result |
|---|---|
| Streaming with four Data workers | 2.1 to 2.5 times more samples per second |
| GPU image augmentation, ACT training | 2.8 times faster on an H100, 3.3 times on an A100 |
| FSDP2 with DCP checkpoints, 8 x H200 node | Checkpoint saves 3 times faster; resume peak memory falls from 143 GB to 42 GB per GPU |
The team also notes that seven image transforms (GaussianNoise, MotionBlur, GaussianPatchBrightness, RandomShadow, CoarseDropout, GammaCorrection and PlanckianJitter) were silent no-ops on uint8 frames before this release. They now apply. If you trained with them enabled in an earlier version, your results never used them. Retrain before you compare runs across versions.
The largest recent public dataset in this format is FineART from Scale AI: 40,000 or more bimanual episodes, 1,718 hours and 1.9 TB. Streaming is aimed at sets of that size.
Which new models does LeRobot v0.7.0 add? #
The release adds two families, according to the Hugging Face blog.
World action models (WAMs): Cosmos 3 (Nano at 16.5B parameters and Edge at 4.6B), FLUX 3 Action from Black Forest Labs, LaWAM from RLinf, and Long-WAM from NVIDIA.
Vision-language-action models (VLAs): G0.5 from OpenGalaxea, LingBot-VLA 2.0 from Robbyant, DM0.5 from Dexmal, Wall-OSS-0.5 from X Square Robot, and FineART-VLA, which adds a language head to π0.5.
Numbers the release blog gives for these models:
| Model | Benchmark | Reported result |
|---|---|---|
| Cosmos 3 Nano / Edge | LIBERO average success | 97.90% / 97.95% |
| LaWAM | LIBERO | 98.5% |
| Long-WAM | LIBERO | 99.00% (IDM), 98.45% (COD) |
| LingBot-VLA 2.0 | RoboTwin 2.0 (clean) | 93.5%, against 82.7% for π0.5 |
LingBot-VLA 2.0 also cuts action chunk latency from about 756 ms to about 120 ms, per the same post. These are numbers the model authors and the LeRobot team report. We did not re-run them. LIBERO scores near 98 to 99 percent are close to saturated, so they say little about real-world skill. For a harder comparison, see our look at RoboColosseum, Physical Intelligence's real-world evaluation.
One license warning stands out. The release blog states: "G0.5's code and weights are released under the non-commercial G0.5 Community License, not Apache-2.0." Check each model card before commercial use. Our post on Verne Robotics v0.5 shows how often "open" robot models hide gated terms.
How does remote inference work? #
lerobot-policy-server runs the model on a GPU machine. The robot's own computer only handles sensors and motors. The release blog says transport uses Zenoh with TLS and access control. To run a rollout against it, use lerobot-rollout --inference.type=remote and install the remote extra.
The old lerobot.async_inference module and the lerobot[async] extra are removed. If your code imports them, it breaks.
This split matters for small robots. A low-cost arm with an embedded board cannot run a multi-billion-parameter VLA. A laptop or a cloud GPU can. The trade-off is network latency and a new attack surface, because a robot that takes commands over a network needs authentication. TLS and access control help, but you still need to set them up.
Simulation got a similar redesign. Every simulator now runs through one vector-env core, in-process or in Docker via lerobot-env-server. The first benchmark on this core is NVIDIA's RoboLab with 120 tabletop tasks. lerobot-eval now reports 95% Wilson confidence intervals, so a score comes with an error bar. That is a quiet but useful change for anyone who compares policies.
What new hardware does it support? #
- YAM: single or bimanual I2RT arms with gravity feed-forward compensation.
- reBot B601: works with both RobStride and Damiao motors.
- Unitree G1: runs NVIDIA's SONIC decoder. It turns 64-dimension motion tokens into 50 Hz joint targets. A 23-DoF variant is also supported.
Feetech and Dynamixel motors now run on one rustypot-backed SerialMotorsBus. The old vendor SDK buses remain as deprecated aliases. Calibration files are unchanged, per the GitHub notes.
Robot imitation learning with a gripper copying a human hand holding a cup, the core idea behind LeRobot policies
For context on humanoid control research that targets the same G1 robot, see SuperMimic from Berkeley and NVIDIA MotionBricks.
What breaks when you upgrade from v0.6.1? #
This is the part to read before you run pip install -U lerobot. The GitHub notes list these changes:
| Area | Change |
|---|---|
| Install | Linux torch and torchvision come from PyPI as CUDA 13 builds. You need NVIDIA driver 580.65 or newer. The cu128 index is removed. macOS needs 14 or newer (15 for MPS). |
| Pins | transformers 5.22, huggingface-hub 1.31 or newer, datasets below 6, Pillow below 14 |
| Tracking | Weights and Biases is replaced by Trackio. --wandb.* becomes--trackio.* . Checkpoints are no longer uploaded as artifacts by default. |
| Distributed training | FSDP1 is replaced by FSDP2 and HSDP. FSDP checkpoints from 0.6.x cannot resume, but model.safetensors still loads. |
| PEFT | --policy.use_peft=true alone now trains LoRA, not a full fine-tune |
| Inference | lerobot.rollout.inference moved tolerobot.inference .select_action may return aPolicyOutput dict. |
| Simulators | lerobot.envs.<sim> moved tolerobot.sim.integrations.<sim> . Configs are renamed*EnvConfig . |
| Precision | dtype is the single precision key. In Python passtorch.bfloat16 , not a string. |
Our advice: pin your current version in a lockfile, create a fresh virtual environment for 0.7.0, and test on a simulator before you touch a real arm. Several of these changes are silent. The use_peft change alone can turn a planned full fine-tune into a LoRA run.
How do you try it? #
- Read the docs at huggingface.co/docs/lerobot/v0.7.0 .
- Install in a new environment. Check your driver with
nvidia-smifirst. You need 580.65 or newer on Linux. - Pull a starting checkpoint. The release blog names
lerobot/Cosmos3-Nano,lerobot/wall-oss-0.5andlerobot/dm05_base. - Test streaming on a public set: add
--dataset.streaming=trueto a training command. - Start an interactive rollout in simulation, then try
/subtaskwith a text-capable policy.
What this means for builders #
If you build with low-cost arms, the remote policy server is the most useful change. It lets a cheap robot use a large model without a large onboard computer. If you train policies, streaming and faster checkpoints cut the hardware bar. If you maintain older code, budget a day for migration. The project is also growing. The team says its Discord and X communities each passed 20,000 members and GitHub is nearing 30,000 stars. We could not verify those community counts independently.
For the wider picture of open robot models released this month, read our coverage of [Ai2's ManiFlow-TB](https://explainx.ai/blog/ai2-maniflow-tb-toyblocks-open-robot-action-model-322m-franka-flow-matching-apache-57-percent-success-2026), the [HIW-500 humanoid dataset](https://explainx.ai/blog/hiw-500-bitrobot-unitree-humanoid-dataset-lerobot-2026) and the [RoboValue benchmark](https://explainx.ai/blog/robovalue-benchmark-robotic-value-models-15-sim-20-real-tasks-lerobot-2026). For NVIDIA's simulation side, see the [Cosmos interactive simulation guide](https://explainx.ai/blog/nvidia-cosmos-interactive-simulation-guide-sim-auto-transfer-auto-h-dreams-nim-vram-table-2026).
If a robot on your network can take text commands, treat it like any agent with real-world reach. [AgentBeam](https://agentbeam.com/?ref=explainx), the agent security platform from the explainx.ai team, stops AI agents before they take dangerous actions.
Related reading #
Version numbers, flags and results are accurate as of November 6, 2026. Check the official release notes for later changes.