LeRobot has released version 0.6.0, introducing a comprehensive suite of tools designed to close the robot learning loop. This update emphasizes world model policies, unified reward systems, streamlined data pipelines, and centralized evaluation, significantly advancing vision-language-action model deployment. The model zoo expands with GR00T N1.7, MolmoAct2, EO-1, Multitask DiT, and EVO1. Optimized for varying hardware constraints, these architectures support real-time inference on modest GPUs and zero-shot deployment on standard robotic arms. The release also introduces three world model policies that train agents to predict future trajectories. By integrating future imagination directly into the training loop while eliminating inference overhead through latent-space prediction and action chunk denoising, these models enhance policy reasoning without sacrificing performance. To address success detection, the framework deploys a unified reward API featuring Robometer and TOPReward. These tools evaluate task progress directly from raw video and language instructions without requiring task-specific fine-tuning. The API natively supports reward-aware behavior cloning and generates per-frame progress metrics, enabling rigorous dataset inspection and automated quality scoring. Data handling receives substantial upgrades, including end-to-end depth sensor recording, VLM-powered automated language annotation, and fully customizable video encoding. Parallel multi-camera decoding and persistent worker caching double data loading speeds, while deterministic sampling ensures fully reproducible training runs across interrupted sessions. Evaluation and deployment workflows are now unified under dedicated command-line interfaces. The lerobot-eval CLI aggregates nine simulation benchmarks, including LIBERO-plus, RoboTwin 2.0, and RoboCasa365, standardizing performance testing across complex manipulation and long-horizon tasks. The new lerobot-rollout CLI streamlines real-robot deployment and introduces a DAgger strategy. This feature records human interventions during policy failures, automatically tagging correction data for immediate fine-tuning and accelerating the robot learning flywheel. Training infrastructure gains native Fully Sharded Data Parallel support for multi-GPU scaling and direct cloud execution via Hugging Face Jobs. Users can deploy distributed environments on-demand, ranging from entry-level to high-end GPUs, effectively eliminating local hardware bottlenecks. Under the hood, the platform improves efficiency with a forty percent reduction in base dependencies and a shift to feature-scoped extras. The codebase enforces strict dependency management, supports mixed-precision training, and integrates seamlessly with Foxglove for remote teleoperation visualization. Compatibility now extends to Wayland, headless systems, and macOS without accessibility overrides. By consolidating training, reward modeling, data curation, and evaluation into a single cohesive workflow, LeRobot 0.6.0 establishes a production-ready foundation for open-source robotic AI, substantially lowering deployment barriers for researchers and developers.