LeRobot v0.6.0 Arrives: Robots That Imagine, Evaluate and Improve Themselves

The open-source robotics framework LeRobot has reached version 0.6.0, delivering one of its most feature-packed updates yet. The release focuses on three ideas captured in its tagline: imagine, evaluate and improve. It introduces policies that learn to predict the future, expands its catalogue of vision-language-action models, adds tools to measure success and makes training and data handling faster and lighter.

Policies That Learn to Imagine

A central question in robotics is whether world models — systems that anticipate what happens next — actually help robot policies perform better. Version 0.6.0 ships three new policies designed to explore that question, each balancing imagination with computational cost.

VLA-JEPA trains a compact model to predict future frames in latent space during learning, then drops the world model at inference so there is no extra runtime cost. Ready-to-use checkpoints, including a pretrained base for fine-tuning, are available on the Hub.

LingBot-VA goes further, predicting future video and actions together chunk by chunk while feeding real observations back in to stay grounded. Users can even save what the robot imagined and compare it to what actually happened, with inference running on a single 24–32 GB GPU. FastWAM pairs a video-generation expert with a compact action expert, learning to dream its own rollouts during training but skipping the dreaming at inference to denoise actions directly.

A Growing Model Zoo

The release expands LeRobot's collection of vision-language-action models (VLAs). The NVIDIA GR00T integration has been upgraded to GR00T N1.7, the newest open generation of the cross-embodiment foundation model, tested for parity against NVIDIA's original implementation. It replaces N1.5, which remains available by pinning an older version.

MolmoAct2 from the Allen Institute for AI is now fully supported, covering fine-tuning, evaluation and real-robot deployment, with checkpoints that can run zero-shot on SO-100/101 hardware. Inference fits in roughly 12 GB, while LoRA fine-tuning works on a single 24 GB GPU.

Other additions include EO-1, built on a Qwen2.5-VL-3B backbone; a Multitask Diffusion Transformer that brings the TRI Large Behavior Models recipe to LeRobot; and EVO1, a lightweight 0.77B-parameter policy small enough to run in real time on modest GPUs.

Knowing When a Robot Succeeds

Success detection and progress estimation have long been missing pieces of the robot learning loop. LeRobot now offers a unified reward models API that mirrors its policies interface, bundling four reward models behind a single system.

Robometer is a general-purpose, pretrained reward model that scores task progress and success from raw video and a language instruction, without task-specific training. It was trained on more than one million robot trajectories. TOPReward takes a fully zero-shot approach, wrapping an off-the-shelf vision-language model to turn any capable model into a reward function. Both come with labeling scripts that write per-frame progress curves into datasets.

Richer Datasets and Faster Loading

Data handling received major upgrades. Recording is no longer locked to a single codec, with new options exposing the full encoding surface and automatic detection of hardware encoders. Depth support arrives too: connecting an Intel RealSense camera lets LeRobot record depth maps end to end across a wide range of supported robots.

Datasets can now store rich language annotations such as timestamped subtasks, plans, corrections and per-camera question-answer pairs, filled in automatically by a new annotation CLI that uses a vision-language model to watch episodes. Training on video datasets is up to roughly twice as fast, with parallel decoding, leaner memory use and deterministic, resumable sampling. Loading a subset of a large dataset dropped from minutes to milliseconds in the team's benchmark.

One CLI to Evaluate Them All

Building on the previous release's push to make LeRobot an evaluation hub, v0.6.0 adds six new simulation benchmarks, all runnable through the same command-line tool. They include LIBERO-plus, which stress-tests policies with roughly 10,000 perturbed task variants; RoboTwin 2.0, covering 50 bimanual manipulation tasks; RoboCasa365, spanning 365 kitchen tasks; RoboCerebra for long-horizon behavior; RoboMME as a memory exam; and VLABench for knowledge and reasoning.

Together with existing benchmarks, that makes nine benchmark families under one roof, with a guide for adding new ones. Evaluation also runs faster, defaulting to asynchronous vectorized environments.

Deployment, Training and a Leaner Codebase

A new rollout CLI turns deployment into its own workflow, including a DAgger strategy that lets operators take over the moment a policy goes wrong, record the correction and hand control back — creating a data-collection flywheel from a single command. Training now supports fully sharded data parallel across multiple GPUs, and the same training command can run in the cloud by adding one flag, with jobs ranging from a single T4 to eight H200 GPUs on a pay-as-you-go basis.

Installation is now genuinely lightweight, with roughly 40% fewer base dependencies and feature-scoped extras for everything else. The release also adds browser-based tooling through LeLab, VR teleoperation via NVIDIA's Isaac Teleop, hardware guidance for newcomers and hundreds of bug fixes.

The LeRobot team credits contributions from academia, industry and hobbyists who chose the platform as home for their models and benchmarks. If you are curious about the future of open-source robotics, this release is worth a closer look — and worth sharing with anyone building the next generation of intelligent machines.

Source: Hugging Face Blog