Hugging Face ships LeRobot v0.6.0 with world-model robot policies that imagine the future before acting

Hugging Face has released version 0.6.0 of LeRobot, its open-source robot-learning library, headlined by a new class of "world model" policies that predict what will happen next before a robot commits to an action, alongside a wave of new vision-language-action models and reward models.

What's new

Hugging Face frames the release around closing a feedback loop: "This new release is about closing the robot learning loop: policies that imagine the future before acting, reward models that tell you when your robot succeeds, a deployment CLI that turns failures into training data, and six new simulation benchmarks to measure it all."

The centerpiece is a set of world-model policies. VLA-JEPA learns to predict future frames in latent space during training, but the world model "disappears at inference, so you get world-model supervision at zero extra inference cost." LingBot-VA is described as an autoregressive video-action model that predicts future video and actions together and runs on a single 24-32 GB GPU. FastWAM pairs a roughly 5-billion-parameter video-generation expert with a compact action expert but skips the prediction step entirely at inference time for speed.

On the vision-language-action side, the release adds GR00T N1.7, described as "the newest open generation of NVIDIA's cross-embodiment foundation model," replacing the prior N1.5. MolmoAct2 is tuned to fit inference in roughly 12 GB at bf16 precision, with LoRA fine-tuning possible on a single 24 GB GPU. EVO1 packs its policy into 0.77 billion parameters, small enough to fine-tune and run in real time.

A new reward-models API ships alongside the policies. Robometer is a pretrained, general-purpose reward model built on a dataset of more than one million robot trajectories. TOPReward wraps an off-the-shelf vision-language model and reads the log-probability of the token "True" as its reward signal.

On the tooling side, the release adds six new simulation benchmarks unified under a lerobot-eval command, a lerobot-rollout CLI supporting DAgger-style human-in-the-loop corrections, FSDP training support, and cloud training via Hugging Face Jobs. Datasets gain depth-sensor support, an automatic language-annotation pipeline, and custom video encoding, with data loading now up to roughly 2x faster.

Context

LeRobot is Hugging Face's push to bring the open, community-driven tooling model it built for language models into robotics, where training stacks, datasets, and evaluation methodology have historically been fragmented across labs and hardware vendors. This release directly tests one of the more contested questions in robot learning: whether giving a policy an internal "world model" to predict outcomes before acting actually improves performance over policies that map perception straight to action.

Why it matters

World models have shown promise in other domains as a way to make agents more sample-efficient and better at planning, and this release is a concrete, reproducible test of that idea for physical robots rather than a research paper on video prediction in the abstract. Because the checkpoints, training code, and new benchmarks all ship together and are runnable on consumer-scale GPUs (many of the new VLAs are explicitly tuned to run in the 12-32 GB range), the release lowers the bar for independent researchers and smaller robotics teams to test world-model approaches themselves rather than relying on results reported by a handful of well-resourced labs.