World Models and DAgger Integration Ship in LeRobot v0.6.0

Hugging Face released LeRobot v0.6.0, transitioning the open-source library from reactive control to closed-loop robot learning. The update introduces world models that predict future states, reward models that measure per-frame task progress, and a dedicated deployment CLI for human-in-the-loop corrections. For developers building robotics policies, this release standardizes success detection and environment simulation into a unified pipeline.

Predictive Policies and World Models

The v0.6.0 release introduces architectural shifts designed to move beyond immediate reactive control. The integration of Joint-Embedding Predictive Architecture (JEPA) via VLA-JEPA allows robots to learn by simulating the future consequences of their actions.

The library adds FastWAM, a specialized world model variant optimized for speed and real-time response latency. For language-driven tasks, LingBot-VA incorporates visual-action planning directly into a world-model-based policy. By introducing predictive mechanisms, the library transitions robot behavior from reactive adjustments to anticipatory actions.

The vision-language-action (VLA) expansion brings several new models to the library. These include GR00T N1.7, which replaces the previous N1.5 version, alongside EO-1, EVO1, and Multitask DiT. MolmoAct2 operates zero-shot on physical robots like the SO-100 and SO-101 using approximately 12GB of VRAM in bf16.

Closed-Loop Evaluation Framework

Robotics engineering requires standardized methods to evaluate AI output during continuous operation. LeRobot introduces a Reward Models API to determine task success or progress per frame, providing a systematic approach to success detection that was previously fragmented in open robotics.

Robometer operates as a 4B parameter model pre-trained on over 1 million trajectories for success detection. TOPReward functions as a zero-shot reward metric utilizing log-probabilities from any capable vision-language model.

The new lerobot-eval CLI unifies policy evaluation across six integrated simulation environments. This testing surface includes LIBERO with 130 defined tasks, Meta-World with 50 tasks, and the NVIDIA IsaacLab Arena.

DAgger Deployment and Infrastructure Scaling

Data collection and correction workflows are now formalized through the lerobot-rollout CLI. This deployment interface supports DAgger-style human-in-the-loop interventions, automatically recording human corrections during robot failures as new training data. Datasets now support depth sensing, end-to-end video encoding with custom codecs, and an automated language annotation pipeline powered by VLMs.

Data loading speeds have increased by up to 2x. The core installation footprint is 40% leaner, achieved by shifting hardware-specific dependencies to feature-scoped extras.

Training infrastructure now supports Fully Sharded Data Parallel (FSDP) for models exceeding single-GPU memory limits. Developers can route workloads directly to Hugging Face Jobs via the --job.target command-line flag, scaling from a single T4 to an 8x H200 cluster. Supported frameworks have been updated to PyTorch 2.7 through 2.11, with CUDA 12.8 wheels pinned for Linux installations using the uv package manager.

If you are configuring autonomous policies, route your local execution failures through the lerobot-rollout interface to automatically populate your correction datasets and close the learning loop.