LeRobot published version v0.6.0 on July 7, 2026, introducing new world model policies, Vision-Language-Action (VLA) models, a reward models application programming interface (API), simulation benchmarks, and command-line interface (CLI) tools. The update includes performance enhancements, such as up to 2x faster data loading and a streamlined installation process.
The new world model policies include VLA-JEPA, FastWAM, and LingBot-VA. VLA-JEPA is built on Qwen3-VL-2B and utilizes a JEPA world model during training to anticipate future frames based on the model's actions, though this world model is not employed during inference. Three VLA-JEPA checkpoints are available on the Hub, including a DROID-pretrained base model engineered for fine-tuning. LingBot-VA is an autoregressive video-action model designed to predict video and actions concurrently, feeding real observations back into the system to maintain prediction accuracy. Users can save predicted video output using the parameter --policy.save_predicted_video=true, and its inference runs on a single 24–32 GB graphics processing unit (GPU). FastWAM incorporates an approximately 5B video-generation expert with a compact action expert within a single network, bypassing future imagination during inference to directly denoise action chunks.
Version v0.6.0 also introduces new Vision-Language-Action models including GR00T N1.7, MolmoAct2, EO-1, EVO1, and Multitask DiT. LeRobot upgraded its NVIDIA GR00T integration to GR00T N1.7, which replaces the previous GR00T N1.5 version. GR00T N1.7 features Cosmos-Reason2-2B, built on Qwen3-VL, as its VLM, along with a flow-matching action head. The LeRobot GR00T N1.7 integration has been parity-tested against NVIDIA's original Isaac-GR00T implementation, and flash-attention is an optional feature. Users can pin lerobot==0.5.1 to access GR00T N1.5. MolmoAct2, a VLA model from the Allen Institute for AI, has been ported into LeRobot with support for fine-tuning, evaluation, and real-robot deployment. Its ready-made checkpoints include calibration correction, and it can operate zero-shot on an SO-100/101 robot. MolmoAct2 inference requires approximately 12 GB of memory at bf16 precision, and LoRA fine-tuning fits on a single 24 GB GPU. EO-1, contributed by one of its paper's authors, is a VLA model pretrained on interleaved vision-text-action data, employing a Qwen2.5-VL-3B backbone with a flow-matching action head. EVO1 has 0.77 billion parameters, using an InternVL3-1B backbone with a flow-matching action head, and supports two-stage fine-tuning and Real-Time Chunking. Multitask DiT is a diffusion transformer policy based on the TRI Large Behavior Models recipe, featuring approximately 450 million parameters and conditioned on CLIP vision and language embeddings. It supports both diffusion and flow-matching objectives.
LeRobot v0.6.0 includes a new unified reward models API, lerobot.rewards, featuring Robometer and TOPReward. Robometer is a pretrained general-purpose reward model that assesses task progress and success from raw video and language instructions without requiring task-specific training. Built on Qwen3-VL-4B, Robometer was trained using trajectory comparisons over a dataset of more than one million robot trajectories. TOPReward operates without reward weights, wrapping an off-the-shelf VLM like Qwen3-VL and reading the log-probability of the token "True" given the trajectory video and task instruction. Both Robometer and TOPReward ship with labeling scripts that write per-frame progress curves into datasets.
The release also ships six new simulation benchmarks, all unified under lerobot-eval, and the lerobot-rollout CLI tool, which includes DAgger-style human-in-the-loop corrections. Support for FSDP training and cloud training on HF Jobs has been added. Datasets now include depth sensing support, automatic language annotation pipelines, and custom video encoding capabilities.
forum Comments (0)
No comments yet. Be the first to comment.