LeRobot 0.6.0 adds world models, VLAs, and unified robot evaluation

LeRobot v0.6.0 is a major robotics release focused on three practical goals: helping robot policies imagine what may happen next, evaluating them in a more consistent way, and improving the training and deployment workflow around them. The release adds several world-model-based policies, expands the vision-language-action model zoo, introduces reward model support, upgrades datasets, and brings new simulation benchmarks under a shared evaluation interface. In plain terms, LeRobot is trying to make it easier to build robots that can learn from vision, language, actions, and feedback without every lab rebuilding the same tooling from scratch.

The biggest theme is world models: robot policies that learn an internal sense of the future while they train. VLA-JEPA, LingBot-VA, and FastWAM all explore the same core question: can a robot act better if it learns to imagine likely future observations and actions? VLA-JEPA uses a compact vision-language-action model based on Qwen3-VL-2B and trains it with a JEPA-style world model that predicts future frames in a hidden representation rather than raw pixels. The important engineering detail is that this extra world-model component is used during training but removed at inference, so the robot gets the benefit of future-prediction supervision without paying extra runtime cost. The release also includes ready-to-use checkpoints, including a DROID-pretrained base intended for fine-tuning.

LingBot-VA takes a more direct route: it predicts future video and actions together, chunk by chunk. It is an autoregressive video-action model, meaning it generates the next part of a sequence from what came before, then uses real observations again to keep those predictions from drifting too far away from reality. LeRobot also exposes an option to save the predicted video, letting developers compare what the robot expected to happen with what actually happened. That makes the model useful not only for control but also for debugging. The source notes that inference can run on a single 24-32 GB GPU, which matters because many robotics teams do not have unlimited hardware budgets.

FastWAM approaches the same problem through a world action model. It combines a roughly 5B-parameter video-generation expert with a smaller action expert in one network, so the system learns rollouts during training. At inference time, however, it does not spend time generating imagined futures. Instead, it directly denoises chunks of robot actions. This continues the release’s practical pattern: use richer prediction during training, but keep deployment efficient enough for real systems.

The release also expands LeRobot’s VLA model zoo. NVIDIA GR00T integration has been upgraded to GR00T N1.7, replacing N1.5 in the current LeRobot line. N1.7 uses Cosmos-Reason2-2B, built on Qwen3-VL, feeding a flow-matching action head. The integration is parity-tested against NVIDIA’s Isaac-GR00T implementation, meaning the same inputs should produce the same outputs. Flash-attention is now optional, making installation simpler, and NVIDIA’s published checkpoints can be loaded directly. For teams that still need N1.5, the source says they should pin LeRobot to version 0.5.1.

MolmoAct2 from the Allen Institute for AI is now available inside LeRobot with fine-tuning, evaluation, and real-robot deployment covered. It supports both full fine-tuning and LoRA, a lighter method that adjusts a smaller number of model parameters. Ready-made checkpoints include calibration correction and can be used zero-shot on SO-100/101 robots. The stated hardware footprint is also practical: inference fits in around 12 GB at bf16 precision, while LoRA fine-tuning fits on one 24 GB GPU. EO-1 also joins the library, using a Qwen2.5-VL-3B backbone and a flow-matching action head, and can be trained through the standard lerobot-train workflow.

LeRobot v0.6.0 adds more than policies. The new reward models API gives developers a way to estimate whether a robot succeeded, using models such as Robometer and TOPReward. This is important because robotics evaluation often depends on messy real-world outcomes rather than simple labels. Reward models can help judge progress, guide training, and compare behavior across tasks. The release also unifies six new simulation benchmarks under lerobot-eval, giving teams a single command-line path for evaluation instead of separate benchmark-specific flows.

Datasets receive several quality-of-life and capability upgrades. The release adds end-to-end depth support, so datasets can include depth information in addition to ordinary images. It introduces an automatic language annotation pipeline, which can help turn robot data into richer training material by adding text descriptions at scale. It also allows custom video encoding, so teams can choose codecs that fit their storage, speed, or quality needs. Data loading is described as up to twice as fast, a meaningful improvement because robotics training often bottlenecks on large video-heavy datasets.

Training and deployment tooling also gets stronger. The new lerobot-rollout CLI gives deployment its own command-line tool and supports DAgger-style human-in-the-loop corrections. DAgger is a training method where a human can correct the robot as it acts, producing better examples from the situations the robot actually encounters. FSDP support allows training models larger than a single GPU would normally handle by splitting work across devices, and cloud training through Hugging Face Jobs gives teams another path when local hardware is not enough.

Overall, LeRobot v0.6.0 looks like a consolidation release for modern robot learning. It brings frontier research ideas such as world models, VLAs, reward models, diffusion-style policies, and flow-matching action heads into a more unified open robotics stack. The practical message is that robotics development is moving from isolated model demos toward reusable infrastructure for training, evaluating, correcting, and deploying embodied AI systems.