Hugging Face releases LeRobot v0.6.0 with world-model policies, reward APIs and a deployment CLI
The open-source robotics toolkit now standardizes the full predict-evaluate-correct-retrain loop, landing the same week a wave of open robot foundation models hit Hugging Face.
Hugging Face released LeRobot v0.6.0 on July 7, 2026, the largest update yet to its open-source robot-learning toolkit, adding world-model policies, a reward-models API, six simulation benchmarks and a deployment-oriented rollout CLI. The release was published as tag v0.6.0 on GitHub (commit 30da8e6), with a breaking-changes summary linked from the release page, according to reasoncore.dev's release forensics and coverage of the Hugging Face announcement.
Context
LeRobot has become the default open stack for imitation learning and vision-language-action (VLA) model development on hobbyist and research robots. The timing of v0.6.0 is notable: it arrived the same week Ant Group's Robbyant open-sourced its LingBot robotics stack and NVIDIA published new robotics-adjacent model artifacts, and LeRobot's own model zoo absorbed several of those releases. Robotics datasets on Hugging Face grew from 1,145 in 2024 to 26,991 in 2025, per the figures cited in release coverage — the substrate this tooling is built to organize.
What changed
According to the release coverage, v0.6.0 adds:
- World-model policies: VLA-JEPA with DROID-pretrained checkpoints, LingBot-VA and FastWAM, letting policies predict environment dynamics rather than only mapping observations to actions.
- VLA zoo additions: NVIDIA's GR00T N1.7 and Cosmos-Reason2-2B, plus MolmoAct2, EO-1 and EVO1.
- A reward-models API exposing Robometer and TOPReward for automated rollout scoring. Robometer was trained on RBM-1M, a dataset of more than one million trajectories across 21 embodiments, and outperformed GVL, VLAC and RoboReward baselines in its paper — a first-party result not yet independently reproduced.
- Six simulation benchmarks under a new `lerobot-eval` harness.
- A `lerobot-rollout` CLI with DAgger-style human corrections, so operators can intervene during rollouts and feed corrections back into training.
- Training and infrastructure upgrades: FSDP training support, HF Jobs cloud training, and an installation footprint roughly 40% lighter.
The release requires PyTorch 2.7 or newer and includes breaking CLI and configuration changes, per reasoncore.dev. A practical migration guide was published on July 13.
Why it matters
Earlier LeRobot versions standardized data collection and policy training. Version 0.6.0 extends that standardization to the rest of the loop: predict with world models, evaluate against benchmarks and reward models, correct with a human in the loop, and retrain. That closes a workflow that previously required stitching together bespoke evaluation scripts and ad-hoc scoring. Because the new components are open and model-agnostic, the release also functions as a neutral integration layer at a moment when robot foundation models are arriving weekly from both US and Chinese labs.
The composition of the new model zoo is itself a signal. Alongside NVIDIA's GR00T N1.7 and Cosmos-Reason2-2B sit LingBot-VA — from Ant Group's Robbyant stack — and EO-1 and EVO1. LeRobot is becoming the place where competing embodied-AI ecosystems meet, with a common data format, common evaluation harnesses and now common reward-model interfaces. For smaller labs and university groups, that interoperability is the practical difference between reproducing a result and rebuilding an entire stack.
Details
The breaking changes are concentrated in the CLI and configuration schema; existing training scripts need migration. The lighter install (~40% reduction) lowers the barrier for edge and lab machines. HF Jobs integration moves training to Hugging Face's cloud, an optional path rather than a requirement — local FSDP training remains supported. The reward-models API is the piece with the clearest research lineage: Robometer's RBM-1M training set (>1M trajectories, 21 embodiments) is among the largest reward-model corpora described publicly, though its benchmark wins are vendor-reported.
The `lerobot-rollout` CLI with DAgger-style human corrections addresses a known pain point in imitation learning: policies that fail silently mid-episode with no clean way to capture expert interventions. By making correction capture a first-class CLI workflow rather than a custom data-collection project, v0.6.0 lowers the cost of the iterate-on-real-hardware loop that separates demo videos from deployed systems. Combined with the six new `lerobot-eval` simulation benchmarks, teams can now screen policies in simulation, score them with a shared reward model, and only then commit robot time — all inside one versioned toolchain.
Limitations and caveats
This report is based on Hugging Face's announcement as relayed by secondary coverage and on third-party release forensics; the announcement post itself was not independently fetched. Robometer's claimed superiority over GVL, VLAC and RoboReward baselines is a first-party paper result awaiting independent reproduction. The ~40% install-size reduction and dataset-growth figures are as reported in release coverage. Users should expect migration work: CLI and config changes are explicitly breaking, and the PyTorch 2.7 minimum may conflict with pinned environments.
Sources
- Let's Data Science — Hugging Face releases LeRobot 0.6 robotics toolkit (2026-07-07)
- reasoncore.dev — LeRobot v0.6.0: world models, reward APIs and a deployment CLI (2026-07-06)
- Physical AI Guide — LeRobot v0.6 practical guide (2026-07-13)
*Update note: This post was last reviewed on 2026-07-22. Breaking-change details should be checked against the GitHub release page before upgrading production pipelines.*
Sources
- Let's Data Science — aggregator
- reasoncore.dev — aggregator
- Physical AI Guide — aggregator
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.