LeRobot v0.6.0: Hugging Face Closes the Robot Learning Loop

LeRobot v0.6.0 landed 17 hours ago on the Hugging Face blog. It is the largest single release of the open-source robot learning framework since launch — and it answers the question the field has been arguing about for two years: do world models actually help robot policies?

The answer, in this release, is yes. Three of them, in three different flavors.

Three World-Model Policies Ship Today

The headline addition is a trio of world-model policies that pair video prediction with action generation:

VLA-JEPA pairs a Qwen3-VL-2B vision-language backbone with JEPA-style latent prediction. The world model drops out at inference — it exists only to shape the action head during training. Covered in the VLA-JEPA paper.
LingBot-VA is autoregressive over video and action tokens. Expect 24 to 32 GB of GPU memory at inference.
FastWAM fuses a roughly 5B video-generation expert with an action expert — closer to a diffusion-style action head than an autoregressive one.

The reason this matters: world-model policies let a robot simulate consequences before committing to actions, which is the path to generalization beyond the trajectories you collected.

Five New VLAs Join the Model Zoo

LeRobot’s model zoo now includes five new vision-language-action models, ranging from a few hundred million to multi-billion parameters:

GR00T N1.7 — NVIDIA. Cosmos-Reason2-2B backbone, flow-matching action head. The follow-up to GR00T N1.
MolmoAct2 — Allen AI. About 12 GB at inference.
EO-1 — details emerging from the model card.
Multitask DiT — about 450M parameters, trained with TRI’s Large Behavior Models recipe.
EVO1 — 0.77B parameters. Runs on modest GPUs.

That last one matters most for the people reading this. A 770M-parameter VLA that runs locally on a single mid-tier consumer GPU is the kind of model that ends up on hobbyist desks, not just lab benches.

Reward Models Without Task-Specific Training

The new lerobot.rewards API ships two reward models that need no task-specific training:

Robometer — a Qwen3-VL-4B reward model trained on over one million trajectory comparisons. Covered in the RSS 2026 paper.
TOPReward — a zero-shot trick that uses the log-probabilities of a vision-language model as the reward signal. No fine-tuning at all.

For anyone who has ever hand-coded a reward function for a manipulation task, the appeal is obvious. No reward engineering. No task-specific dataset. Just a number that says “this trajectory is better.”

Six Benchmarks Under One CLI

lerobot-eval is now a unified benchmark runner. Six simulation suites ship today:

LIBERO-plus — roughly 10,000 perturbed variants of LIBERO across 7 axes.
RoboTwin 2.0 — 50 bimanual tasks, 100,000+ trajectories.
RoboCasa365 — 365 kitchen tasks across 2,500 kitchen layouts.
RoboCerebra — long-horizon, language-grounded tasks, 6,660 episodes.
RoboMME — memory benchmarks.
VLABench — vision-language-action benchmarking.

Having these unified under one CLI is a bigger deal than it sounds. Reproducible robot learning requires comparable evaluations, and the field has been running benchmarks across half a dozen different harnesses until now.

Training and Deployment: The Plumbing Ships

Three under-the-hood upgrades matter as much as the headline models:

FSDP training — train models larger than a single GPU’s memory.
Cloud training on Hugging Face Jobs — no local GPU required.
lerobot-rollout CLI — DAgger-style human-in-the-loop corrections. The robot runs a policy, the operator corrects it, and the corrected trajectories go back into the training set.

The DAgger integration is the one to watch. It turns LeRobot from a training framework into a deployment framework that improves itself in the field.

Data Pipeline: 2x Faster, Depth-Aware

Two data-pipeline changes deserve mention:

Data loading is roughly 2x faster — the documented example drops from 275 seconds to 0.06 seconds for subset loads.
Intel RealSense depth is supported end-to-end. Depth is captured in millimeters and compressed as 12-bit video streams.
A custom video codec uses NVENC, VideoToolbox, VAAPI, or QSV depending on the hardware — automatically selected.

The depth support closes a real gap. Most VLA training has been RGB-only. Millimeter-precision depth opens up manipulation tasks that RGB cannot disambiguate — bin picking, transparent objects, dense clutter.

What You Can Actually Do Today

If you have been waiting for a reason to take embodied AI seriously, this release is the reason:

Local training on consumer hardware. EVO1 (0.77B) and Multitask DiT (~450M) fit on a single mid-range GPU. The 2x data-loading speedup makes iteration loops tighter.
Cloud training without infrastructure. HF Jobs handles the heavy lifting. Push a config, walk away, pull the checkpoint.
In-the-loop correction. lerobot-rollout is the first CLI that lets a non-expert teach a policy by correcting it in real time.
Honest evaluation. Run any of the six benchmarks on your trained policy. Compare to published numbers without rebuilding a harness.

The robot learning loop is now closed inside a single open-source framework. Train a policy. Roll it out. Correct it. Re-train. Evaluate against published benchmarks. Repeat — without ever leaving LeRobot.

That is what the v0.6.0 release actually delivers.

Sources:

[1] Remi Cadene, Simon Alibert, et al. “LeRobot v0.6.0: Imagine, Evaluate, Improve.” Hugging Face Blog, published 2026-07-07. https://huggingface.co/blog/lerobot-release-v060

[2] Cadene et al. “VLA-JEPA: Video-Language-Action with JEPA World Modeling.” arXiv:2602.10098, 2026. https://arxiv.org/abs/2602.10098

[3] “Robometer: A Generalist Reward Model for Robotics.” Robotics: Science and Systems (RSS) 2026, arXiv:2603.02115. https://arxiv.org/abs/2603.02115

[4] TechAppleGlobal. “LeRobot v0.6.0 Closing the Robot Learning Loop.” Published 2026-07-07. https://global.techapple.com/2026/07/lerobot-v0-6-0-closing-the-robot-learning-loop/