LeRobot v0.6.0 from Hugging Face: robots learn to predict the future and evaluate themselves

Hugging Face released LeRobot v0.6.0 on July 7, 2026 — a major update to the open-source framework for robot training that closes the loop: robots can now predict action outcomes, automatically evaluate their success, and receive real-time corrections from an operator.

What the new world-model policies can do

World-model policies are the key innovation of this release. Three new architectures enable robots to "imagine" the future directly during training.

VLA-JEPA trains a compact vision-language-action model to predict future frames in latent space. During inference, the world-model is discarded, so speed is unaffected. Three ready-made checkpoints are available, including those pretrained on the DROID dataset.

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

FastWAM combines a ~5B-parameter video generation expert with a compact action expert. During inference, it directly denoises action chunks — without imagination, faster than LingBot-VA.

The release also includes a full "zoo" of VLA models: GR00T N1.7 from NVIDIA with Cosmos-Reason2-2B, MolmoAct2 from 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 real-time chunking support.

How robots now evaluate task success

New reward models allow robots to understand whether they completed a task, without manual annotation.

Robometer — a universal task progress evaluator based on Qwen3-VL-4B, trained on over a million robotic trajectories. It takes video and language instructions as input and returns a success score without fine-tuning for specific tasks.

TOPReward requires no special weights: it wraps ready-made VLM models and reads logarithmic probabilities of success tokens based on video trajectory.

Key facts about the release:

• Release date — July 7, 2026
• Three world-model policies: VLA-JEPA, LingBot-VA, FastWAM
• Robometer trained on 1+ million robotic trajectories (Qwen3-VL-4B)
• Six new simulation benchmarks; nine families total
• Dataset subset loading — from 275 seconds to 0.06 seconds
• Base installation 40% lighter thanks to modular extras

What changed for developers

Six new simulations are unified under the CLI `lerobot-eval`: LIBERO-plus with ~10,000 variations across seven perturbation axes, RoboTwin 2.0 with 50 dual-arm tasks and 100,000+ trajectories, RoboCasa365 with 365 kitchen tasks in 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 DAgger strategy: the operator takes control, corrections are recorded and tagged for the next fine-tuning cycle. Control handoff is smooth.

FSDP support via Accelerate allows training models that don't 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 training loop means when the policy can imagine outcomes, the system can evaluate success, and a human can effectively correct errors," — from the official

LeRobot v0.6.0 release notes.

What this means

LeRobot v0.6.0 offers an open infrastructure for the complete robotic ML cycle: data generation, training with world-models, automatic evaluation through reward models, and human-in-the-loop corrections — all from a single set of CLIs. This lowers the barrier to entry and accelerates iterations without closed proprietary stacks.

Frequently asked questions

How much video memory is needed to run LingBot-VA?

LingBot-VA requires a GPU with 24–32 GB VRAM. If VRAM is less — MolmoAct2 is suitable, which fits in approximately 12 GB when working in bf16 format on SO-100/SO-101.

What is Robometer and how does it evaluate tasks?

Robometer is a reward model based on Qwen3-VL-4B, trained on over a million robotic trajectories. It takes a video episode and text instruction as input and returns a progress score without fine-tuning for specific tasks.