LeRobot v0.6.0: Closing the Robot Learning Loop

Hugging Face has released LeRobot v0.6.0, the largest update to the open-source robotics learning framework to date. The new version centers on three core themes — imagine, evaluate, and improve — introducing world model policies that predict future states, a unified reward model API, and six new simulation benchmarks under a single evaluation interface.

World Models: Policies That Imagine the Future

The headline addition in v0.6.0 is three world model policies, each learning to predict the future during training while taking a different approach to keeping that prediction affordable at inference time.

VLA-JEPA builds a compact vision-language-action model on top of Qwen3-VL-2B. During training, a JEPA world model anticipates upcoming frames from the model’s own actions. At inference, the world model is removed entirely, delivering world-model supervision at zero extra cost. Three pre-trained checkpoints are available on Hugging Face, including a DROID-pretrained base for fine-tuning.

LingBot-VA takes a further step with an autoregressive video-action model that predicts future video and actions together, chunk by chunk, feeding real observations back to keep predictions grounded. Researchers can save what the robot imagined and compare it with actual outcomes. Inference runs on a single 24–32 GB GPU.

FastWAM asks a fundamental question: do world action models actually need test-time future imagination? It pairs a roughly 5-billion-parameter video generation expert with a compact action expert in one network. At inference, it skips the imagination step entirely and directly denoises action chunks.

The VLA Zoo Keeps Growing

NVIDIA’s GR00T has been upgraded to N1.7, replacing the previous VLM with Cosmos-Reason2-2B built on Qwen3-VL, feeding a flow-matching action head. The LeRobot integration has been parity-tested against NVIDIA’s original Isaac-GR00T implementation.

Allen AI’s MolmoAct2 has been fully ported into LeRobot with fine-tuning, evaluation, and real-robot deployment covered end to end. Pre-trained checkpoints with calibration correction support zero-shot deployment on SO-100/101 robots at roughly 12 GB of VRAM. Also joining the lineup are EO-1 (Qwen2.5-VL-3B backbone), the ~450M-parameter Multitask DiT, and the compact 0.77B-parameter EVO1.

Reward Models: Knowing When Your Robot Succeeds

Success detection and progress estimation have long been missing pieces in robot learning. v0.6.0 introduces a unified reward models API with two notable additions: Robometer and TOPReward.

Robometer is a pre-trained, general-purpose reward model built on Qwen3-VL-4B, trained via trajectory comparisons over a dataset of more than one million robot trajectories. Point it at any LeRobot dataset and it scores task progress and success from raw video plus a language instruction, with no task-specific training required.

TOPReward goes fully zero-shot with no reward weights at all. It wraps an off-the-shelf VLM and reads the log-probability of the “True” token given the trajectory video and the task instruction, turning any capable VLM into a reward function.

Data, Training, and Benchmark Infrastructure

On the data side, LeRobot now supports depth sensing end to end with Intel RealSense cameras, compresses depth maps as 12-bit video streams, and offers an automatic language annotation pipeline using VLMs. Video data loading is up to roughly 2x faster, with multi-camera frames decoding in parallel.

Training gains FSDP (Fully Sharded Data Parallel) support through Accelerate, enabling models that exceed single-GPU memory. Cloud training through Hugging Face Jobs works by adding a single flag — from a T4 to 8x H200, all accessible with the same lerobot-train command.

Six new simulation benchmarks run through the unified lerobot-eval CLI: LIBERO-plus stress-tests with roughly 10,000 perturbed variants, RoboTwin 2.0 covers 50 bimanual tasks, RoboCasa365 spans 365 kitchen tasks across procedurally generated environments, RoboCerebra evaluates long-horizon behavior, RoboMME tests memory capabilities, and VLABench examines knowledge and reasoning in manipulation.

Deployment gets its own CLI in lerobot-rollout, with the DAgger strategy enabling researchers to watch their policy run, take over with a leader arm when it fails, and feed every correction back into the next fine-tuning cycle — closing the robot learning loop.

More details available on the Hugging Face Blog.