LeRobot v0.6.0 adds world models, new VLAs, and reward APIs for robot learning

LeRobot v0.6.0 closes the robot learning loop with policies that imagine the future before acting, reward models that detect success, and a deployment CLI that turns failures into training data. Three new world model policies debut: VLA-JEPA (built on Qwen3-VL-2B) predicts future frames in latent space during training and discards the world model at inference for zero extra cost; LingBot-VA autoregressively predicts video and actions together, feeding real observations back to stay grounded; FastWAM pairs a ~5B video-generation expert with a compact action expert, learning to dream rollouts then skipping the dream at inference. All are fine-tunable via lerobot-train.

The release also grows the VLA zoo with GR00T N1.7 (NVIDIA's latest cross-embodiment model using Cosmos-Reason2-2B), MolmoAct2, EO-1, EVO1, and Multitask DiT. A new reward models API includes Robometer and TOPReward. Six simulation benchmarks unify under lerobot-eval, and the lerobot-rollout CLI supports DAgger-style human‑in‑the‑loop corrections. Datasets now support depth, automatic language annotations (via VLMs), custom video codecs, and up to 2x faster loading. FSDP enables training bigger models, and HF Jobs provides cloud training. The codebase is leaner with optional flash-attention and streamlined installs.