Hugging Face released LeRobot version 0.6.0 on July 7, 2026, introducing world model policies, reward models, and six new simulation benchmarks to its open-source robotics framework. The update expands the platform's model library, adds depth sensing, and cuts base installation dependencies by roughly 40 percent.
Three new world model policies predict future observations during training. VLA-JEPA uses a Qwen3-VL-2B backbone to predict latent space representations while acting, discarding the world model at inference to avoid extra computation. LingBot-VA predicts future video frames and actions together on a single 24 to 32 gigabyte GPU. FastWAM pairs a 5-billion-parameter video generation expert with an action expert, skipping video generation during inference to denoise action chunks directly.
Model library and reward APIs
Hugging Face also expanded its vision-language-action model integrations. The framework upgrades its NVIDIA integration to GR00T N1.7, replacing the prior vision-language model with Cosmos-Reason2-2B linked to a flow-matching action head. Additional additions include MolmoAct2 from the Allen Institute for AI, the 3-billion-parameter EO-1 model, the 450-million-parameter Multitask Diffusion Transformer, and EVO1, a 0.77-billion-parameter policy built on InternVL3-1B.
A unified reward models interface under lerobot.rewards tracks task progress and success. The API incorporates Robometer, a 4-billion-parameter model built on Qwen3-VL-4B that scores progress across raw video feeds using trajectory data from over one million episodes. It also includes TOPReward, which measures task success zero-shot by evaluating the log-probability of positive completion tokens using standard vision-language models.
Benchmarks and deployment CLI
Six simulation benchmarks join the framework through the lerobot-eval CLI: LIBERO-plus, RoboTwin 2.0, RoboCasa365, RoboCerebra, RoboMME, and VLABench. These additions bring the framework's total benchmark count to nine families, paired with async vectorized environments that speed up parallel evaluations by up to two times.
A dedicated deployment tool, lerobot-rollout, introduces five operational strategies for physical hardware. Its DAgger workflow allows operators to pause active policy execution, take control using a leader arm with pose alignment to eliminate motion jerks, and save tagged correction frames directly into new training datasets.
Data pipelines and cloud training
Dataset tools in the release introduce 12-bit depth map recording via Intel RealSense hardware, alongside an automated language annotation pipeline using models like Qwen2.5-VL-7B-Instruct. Data loading speeds have increased by up to two times through parallel frame decoding and uint8 process transfers, while dataset subset loading dropped from minutes to milliseconds in benchmarks.
Multi-GPU training now supports Fully Sharded Data Parallel execution through Accelerate, alongside direct cloud job submission on HF Jobs across hardware ranging from single T4 GPUs to eight NVIDIA H200s. The release also adds LeLab, a browser UI for SO-ARM101 workflows, Isaac Teleop support for controlling SO-101 robots using virtual reality controllers over CloudXR or OpenXR, and baseline support for PyTorch versions 2.7 through 2.11.
