LeRobot v0.6.0 imagines the future during training, then makes inference free

The central tension in robot learning has always been the gap between what a policy sees and what it needs to predict. A robot grasping a cup doesn't just need to know the current joint angles. It needs to anticipate where the cup will be by the time its fingers close. Traditional approaches either ignore that future entirely (behavioral cloning) or simulate it at crippling compute cost (model-predictive control with video generation).

World models that train hard, infer light

LeRobot v0.6.0 attacks this trade-off head-on with three world model policies that learn to imagine the future during training but shed that imagination at inference. The most striking design is VLA-JEPA, which pairs a compact VLA built on Qwen3-VL-2B with a JEPA world model that must anticipate upcoming frames from the model's own actions. Because the world model disappears at inference, the policy gets the benefit of future-awareness during training at zero extra runtime cost. Three ready-to-use checkpoints are on the Hub, including a DROID-pretrained base for fine-tuning.

LingBot-VA takes a different route: an autoregressive video-action model that predicts future video and actions together, chunk by chunk, using real observations to keep its imagination grounded. Users can even save what the robot imagined and compare it with actual outcomes. The model runs on a single 24-32 GB GPU. FastWAM, meanwhile, pairs a ~5B video-generation expert with a compact action expert in a single network. It learns to dream its own rollouts before skipping the dream at inference.

The VLA zoo gets five new residents

The release also ships a wave of new vision-language-action models, reflecting the accelerating pace of VLA research. NVIDIA's GR00T integration jumps to N1.7, swapping the previous VLM for Cosmos-Reason2-2B with an Isaac-GR00T parity-tested implementation. MolmoAct2 from the Allen Institute for AI is now fully supported with fine-tuning, evaluation, and real-robot deployment, requiring only ~12 GB at bf16. EO-1, contributed by one of its own authors, brings a Qwen2.5-VL-3B backbone with flow-matching. The Multitask Diffusion Transformer lets a single ~450M-parameter model learn many tasks conditioned on natural language. EVO1 packs a 0.77B-parameter InternVL3-1B design for modest GPU budgets.

Reward models: finally knowing when the robot succeeds

Success detection has long been a missing piece in the robot learning loop. LeRobot v0.6.0 fills it with a unified reward models API behind the lerobot.rewards interface. Robometer, a pretrained general-purpose reward model built on Qwen3-VL-4B, scores task progress and success from raw video plus a language instruction with no task-specific training. It was trained on more than one million robot trajectories. TOPReward goes fully zero-shot: it wraps any VLM and reads the log-probability of the token "True" given trajectory video and instruction. Both ship with labeling scripts that write per-frame progress curves into datasets, enabling reward-aware behavior cloning and dataset quality inspection.

Benchmarks: one CLI to rule them all

The release folds six new simulation benchmarks into a single lerobot-eval CLI. LIBERO-plus stress-tests VLAs with roughly 10,000 perturbed task variants across seven axes. RoboTwin 2.0 covers 50 bimanual manipulation tasks with heavy domain randomization. RoboCasa365 spans 365 kitchen tasks in 2,500 procedurally generated kitchens. RoboCerebra tests long-horizon behavior with chained sub-goals. RoboMME is a memory exam covering counting, hidden objects, and imitation. VLABench tests knowledge and reasoning in manipulation, from physics questions to composite tasks like brewing coffee. Each ships with a Docker image and a SmolVLA baseline checkpoint tested in CI.

Leaner code, faster data, cloud training

The codebase itself gets leaner: pip install lerobot now carries roughly 40% fewer base dependencies with feature-scoped extras. Video encoding exposes the full encoder surface with hardware acceleration probing for NVENC, VAAPI, and others. Depth recording is end-to-end with Intel RealSense support. Data loading accelerates up to 2x with multi-camera parallel decoding and persistent worker caches. The new lerobot-annotate CLI uses a VLM to automatically timestamp subtasks, plans, and corrections across episodes. FSDP training lets models larger than a single GPU be sharded across workers. And the same lerobot-train command runs in the cloud with a single , job.target flag, supporting anything from a T4 to 8x H200.

A DRY approach to robot learning

The guiding philosophy behind v0.6.0 is Don't Repeat Yourself, applied not to code but to compute. Every new feature optimizes for moving intelligence from runtime to training time: world models that vanish at inference, reward models that pretrain on a million trajectories so you do not have to re-label, benchmarks that test many perturbations without retraining. The result is a framework that does not just collect models. It systematizes the growing understanding that robot learning's bottleneck is not model architecture but data efficiency and evaluation rigor.

Together with the nine benchmark families now unified under one roof, the new lerobot-rollout CLI supporting DAgger-style corrections, and cloud training on HF Jobs, LeRobot v0.6.0 positions itself as the operating system for open-source robotics research. The robot learning flywheel, deploy, collect corrections, fine-tune, repeat, is now a CLI flag away.