LeRobot v0.6.0: World Model-Based Policies, New VLA, Benchmarks and CLI for Evaluation

LeRobot v0.6.0 introduces three peaceful model policies: VLA-JEPA (future prediction in latent space without inference cost), LingBot-VA (autoregressive video and action prediction with the ability to save predicted videos), FastWAM (5 billion parameter video generator, bypassing "dreaming" at inference). The VLA zoo has been expanded: GR00T N1.7 (NVIDIA, Cosmos-Reason2-2B with flow-matching), MolmoAct2 (Allen Institute for AI, support for full training and deployment cycle, ~12 GB at inference, 24 GB for LoRA), EO-1 (Qwen2.5-VL-3B with flow-matching), Multitask DiT (450M parameters, conditioned on CLIP), EVO1 (0.77 billion parameters, InternVL3-1B). A unified API has been introduced for reward models: Robometer (Qwen3-VL-4B, pretrained general model, evaluates progress and success) and TOPReward (zero-shot setup, uses log-probability of the "True" token). Six new simulation benchmarks: LIBERO-plus (10,000 perturbed variants), RoboTwin 2.0 (50 tasks on SAPIEN with domain randomization), RoboCasa365 (365 tasks across 2,500 kitchens), RoboCerebra (long horizons, 3–6 subgoals), RoboMME (memory tests), VLABench (knowledge and reasoning). All benchmarks can be run via the unified CLI lerobot-eval. The CLI lerobot-rollout includes strategies: sentry (continuous recording with rotation), highlight (ring buffer), episodic (classical recording), and dagger (human corrections with intervention flag). In datasets: video encoding with any codec (including hardware NVENC, VideoToolbox, VAAPI, QSV), support for depth from Intel RealSense, automatic language annotation via VLM (lerobot-annotate), up to 2x data loading speedup through parallel decoding and compact uint8 frames. FSDP training and cloud training on HF Jobs.