Hugging Face's LeRobot v0.6.0: Robots Learn to Predict the Future and Grade Themselves

On July 7, 2026, Hugging Face released LeRobot v0.6.0, a major update to its open-source robot-learning framework that closes the loop: robots can now predict the outcomes of their actions, automatically score their own success and receive corrections from an operator in real time.

What the new world-model policies can do

World-model policies are the key addition in this release. Three new architectures let a robot "imagine" the future during training.

VLA-JEPA trains a compact vision-language-action model to predict future frames in latent space. At inference time the world model is discarded, so speed does not suffer. Three ready-made checkpoints are available, including ones pretrained on the DROID dataset.

LingBot-VA works as an autoregressive video-action model: it predicts frames and actions sequentially and uses real observations for self-verification. Inference fits on a single GPU with 24–32 GB of VRAM.

FastWAM pairs a ~5B-parameter video-generation expert with a compact action expert. At inference it denoises action chunks directly, without imagination, and runs faster than LingBot-VA.

The release also ships a whole "zoo" of VLA models: NVIDIA's GR00T N1.7 with Cosmos-Reason2-2B, MolmoAct2 from the Allen Institute (SO-100/101, ~12 GB in bf16), EO-1 with Qwen2.5-VL-3B, Multitask DiT (~450M parameters) and the compact EVO1 (0.77B) with support for real-time chunking.

How robots now judge task success

The new reward models let a robot tell whether it completed a task, with no manual labeling.

Robometer is a general-purpose task-progress estimator built on Qwen3-VL-4B and trained on more than a million robot trajectories. It takes a video and a language instruction and returns a success score without fine-tuning for the specific task.

TOPReward needs no special weights: it wraps off-the-shelf VLMs and reads the log probabilities of success tokens based on the trajectory video.

Key facts about the release:

• Release date: July 7, 2026
• Three world-model policies: VLA-JEPA, LingBot-VA, FastWAM
• Robometer was trained on 1+ million robot trajectories (Qwen3-VL-4B)
• Six new simulation benchmarks; 9 families in total
• Loading a dataset subset drops from 275 seconds to 0.06 seconds
• The base installation is 40% lighter thanks to modular extras

What changed for developers

Six new simulations are unified under the `lerobot-eval` CLI: LIBERO-plus with ~10,000 variants across seven perturbation axes, RoboTwin 2.0 with 50 bimanual tasks and 100,000+ trajectories, RoboCasa365 with 365 kitchen tasks across 2,500 procedurally generated kitchens, RoboCerebra with long-horizon episodes of 3–6 subtasks, RoboMME for testing memory and object counting, and VLABench for knowledge and reasoning.

The `lerobot-rollout` CLI includes a DAgger strategy: the operator takes over control, and the corrections are recorded and tagged for the next fine-tuning cycle. The handover happens without jerks.

FSDP support through Accelerate makes it possible to train models that do not fit on a single GPU: parameters, gradients and optimizer state are sharded across accelerators, and the checkpoint is assembled into a single `model.safetensors`.

"Closing the robot-learning loop means a policy that can imagine outcomes, a system that can evaluate success, and a human who can correct mistakes efficiently," reads the official

LeRobot v0.6.0 release.

What this means

LeRobot v0.6.0 offers open infrastructure for the full robotics ML cycle: data generation, training with world models, automatic evaluation through reward models and human-in-the-loop corrections, all from one set of CLIs. This lowers the barrier to entry and speeds up iteration without closed, proprietary stacks.

Frequently asked questions

How much video memory does LingBot-VA need?

LingBot-VA requires a GPU with 24–32 GB of VRAM. If you have less, MolmoAct2 is an option: it fits in roughly 12 GB in bf16 format on SO-100/SO-101.

What is Robometer and how does it score tasks?

Robometer is a reward model built on Qwen3-VL-4B and trained on more than a million robot trajectories. It takes an episode video and a text instruction and returns a progress score without fine-tuning for the specific task.