本文由 AI 分析生成
建立時間： 2026-04-05 來源： https://arxiv.org/html/2602.22818v1
Summary
The LeRobot paper formalizes the library’s four-pillar architecture: unified robot API, standardized LeRobotDataset format, optimized async inference stack, and clean PyTorch algorithm implementations. Key community metric: 16,000+ datasets from 2,200+ contributors as of September 2025, with 3.9M+ episodes — exceeding what any centralized team could collect. The async inference design (decoupling action prediction from execution) is the most technically interesting architectural decision, enabling large models to run on remote servers while the robot runs at real-time control frequency locally.
論文形式化了 LeRobot 的四柱架構：統一機器人 API、標準化 LeRobotDataset 格式、優化非同步推論堆疊，以及乾淨的 PyTorch 演算法實作。截至 2025 年 9 月有 16,000+ 資料集來自 2,200+ 貢獻者，3.9M+ 個 episode。
Key Points
- SO-100/101 cost: ~€225 — dramatically lowers hardware entry barrier
- StreamingLeRobotDataset: on-demand frame fetching, bounded memory regardless of dataset size
- Async inference architecture: policy runs on remote GPU server; robot receives action stream locally — decouples compute from control loop
- ACT: 52M params, ~5ms inference on RTX 4090 — usable on embedded hardware
- π₀: 3.5B params — requires server-class GPU; async inference makes it viable for real robots
- SERL / TD-MPC: reinforcement learning implementations available alongside behavioral cloning
Insights
The async inference architecture is underappreciated. It means a €225 SO-100 arm can run a 3.5B parameter VLA policy — the compute lives elsewhere. This is the same pattern as cloud-robotics, but implemented at the framework level.
The community data scale (16K+ datasets) is the real moat. Pretraining SmolVLA on 481 of these datasets lifted SO-100 task success from 51.7% to 78.3% — the data flywheel is working.
非同步推論架構意義重大：一支 €225 的 SO-100 機械臂可以跑 3.5B 參數的 VLA 策略，計算資源在遠端。社群資料規模（16K+ 資料集）是真正的護城河，預訓練 SmolVLA 將成功率從 51.7% 提升到 78.3%。
Connections
- Clippings-lerobot-v040-release-supercharging-oss-robot-learning
- Clippings-lerobot-smolvla-lightweight-foundation-model
Raw Excerpt
As of September 2025, 16,000+ datasets from 2,200+ individual contributors are openly shared via the LeRobotDataset format.