LeRobot v0.4.0: Major OSS Robotics Release — Datasets v3.0, Plugin System, New VLA Models

本文由 AI 分析生成

建立時間： 2026-04-05 來源： https://huggingface.co/blog/lerobot-release-v040

Summary

LeRobot v0.4.0 is the most significant release since launch. It delivers three structural upgrades: (1) Datasets v3.0 with chunked episodes enabling streaming of >400GB collections like OXE without full download; (2) a plugin system that makes adding third-party hardware a pip install rather than a core library PR; (3) integration of two powerful VLA foundation models — Physical Intelligence’s π0.5 (4B) and NVIDIA’s GR00T N1.5 (3B). On the tooling side, multi-GPU training via Accelerate and a modular Processors pipeline round out a release that shifts LeRobot from research toy to production-capable framework.

LeRobot v0.4.0 是自發佈以來最重要的版本。三大結構性升級：(1) Datasets v3.0 支援分塊 episode 格式，可串流 >400GB 資料集；(2) 插件系統讓新增硬體從修改核心庫變成 pip install；(3) 整合兩個強大 VLA 基礎模型 — Physical Intelligence 的 π0.5（4B）和 NVIDIA 的 GR00T N1.5（3B）。

Key Points

• Datasets v3.0: chunked Parquet + MP4 format; lerobot-edit-dataset CLI for merge/split/delete/feature-add
• Plugin system: pip install lerobot_teleoperator_xyz pattern; Reachy 2 and phone teleoperation available
• π0.5 (4B): open-world generalization via co-training on heterogeneous web + robot data
• GR00T N1.5 (3B): cross-embodiment model; post-trainable directly via LeRobot; trained on real + synthetic + internet video
• Processors Pipeline: PolicyProcessorPipeline (batched tensors) + RobotProcessorPipeline (real-time control) — clean separation of concerns
• Multi-GPU: accelerate launch --multi_gpu — 50% time reduction with 2 GPUs
• Simulations: LIBERO (130+ VLA tasks) and Meta-World (50+ tasks) integrated

Insights

The plugin system is architecturally significant: it externalizes hardware complexity without fragmenting the ecosystem. This mirrors how Hugging Face handled model loading — a standard interface, community implements adapters. If the pattern holds, LeRobot’s hardware support surface will grow faster than the core team can maintain manually.

The simultaneous integration of π0.5 and GR00T N1.5 signals that LeRobot is positioning itself as the universal evaluation and fine-tuning substrate for open VLA models — not just a training library.

插件系統在架構上意義重大：將硬體複雜度外部化而不分裂生態系統，類似 HuggingFace 處理模型載入的方式。同時整合 π0.5 和 GR00T N1.5 表明 LeRobot 正將自己定位為開放 VLA 模型的通用評估和微調平台。

Connections

• Clippings-vla-0-building-state-of-the-art-vlas-with-zero-modification
• libero-benchmarking-knowledge-transfer-lifelong-robot-learning

Raw Excerpt

A brand-new plugin system to revolutionize how you integrate third-party hardware with LeRobot, making connecting any robot, camera, or teleoperator as simply as a pip install.