LeRobot v0.6.0 teaches robots to dream with 3 world models that vanish at runtime

LeRobot v0.6.0 just dropped, and the team is betting big on a single question: do world models actually help robot policies? The new release ships three different policies—VLA-JEPA, FastWAM, and LingBot-VA—each designed to imagine future states during training. It’s a clever way to get the benefits of planning without the crippling inference costs that usually come with it. The standout trick is in VLA-JEPA, which uses a latent-space world model to supervise a compact VLA built on Qwen3-VL-2B. That world model disappears when you deploy, leaving you with a policy that learned from imagined futures but runs at zero extra cost. FastWAM takes a similar gamble, pairing a ~5B video-generation expert with a compact action expert and then ditching the dreaming entirely at test time in favor of directly denoising action chunks.

The model zoo is also swelling. NVIDIA’s GR00T N1.7 enters the chat, swapping its previous VLM for Cosmos-Reason2-2B with a flow-matching action head. The LeRobot integration is parity-tested against NVIDIA’s original implementation, so you can load published checkpoints directly. MolmoAct2 from the Allen Institute for AI is also ported in with a full lifecycle: fine-tuning, eval, and deployment on an SO-100 arm, all fitting in about 12 GB of VRAM. On the smaller side, the Multitask Diffusion Transformer brings TRI’s Large Behavior Models recipe to a train-it-yourself ~450M-parameter diffusion model.

Infrastructure got a serious polish. A new reward models API—with Robometer and TOPReward—lets you score rollouts programmatically. Six simulation benchmarks are now unified under a single lerobot-eval command, and the lerobot-rollout CLI introduces DAgger-style human-in-the-loop corrections. On the data side, loading is up to 2x faster, and there’s a new automatic language annotation pipeline. The codebase itself is leaner, with FSDP training and cloud training on Hugging Face Jobs now supported.

What’s genuinely interesting here is the push to make training-time imagination practical. World models have been an attractive idea for robotics for years, but the compute bill always killed the vibe. These policies are trying to split the difference: learn from imagined futures, then throw the expensive parts away. Whether the resulting behaviors are actually better than a well-tuned baseline is the real test, but the approach is at least honest about the cost constraints that make or break real-world deployment.