本文由 AI 分析生成
Summary
USC + NVIDIA present AutoMate: a framework for training robotic assembly policies in simulation that transfer directly to real hardware without additional fine-tuning, across 100 diverse part geometries. Key contributions include an “assembly-by-disassembly” demonstration collection method, a three-stage generalist policy training pipeline (BC → DAgger → Curriculum RL), and a sim-to-real gap of only 4.2% on Franka Panda hardware.
USC 與 NVIDIA 合作的 AutoMate 框架，在 100 種幾何形狀裝配任務中達到 86.5% 真實世界成功率，Sim-to-Real Gap 僅 4.2%。核心貢獻：「拆解反演示範」繞過手動收集裝配示範的困難，以及三階段 Generalist Policy 訓練（BC→DAgger→Curriculum RL）。
Prerequisites
- Assembly-by-Disassembly — collect disassembly demonstrations, reverse them for assembly training data
- DAgger — Dataset Aggregation: interactive data collection where expert corrects agent mistakes
- Dynamic Time Warping (DTW) — for matching imitation segments in RL reward during training
Core Idea
Training a generalist policy across 100 diverse assembly geometries by: (1) avoiding the hard problem of collecting assembly demonstrations via disassembly reversal, (2) distilling specialist policies into a generalist via BC+DAgger+CurriculumRL, and (3) using domain randomization + imitation-augmented RL to close the sim-to-real gap.
Results
| Policy Type | Simulation | Real World | Sim-to-Real Gap |
|---|---|---|---|
| Specialist (80 parts) | ~80% | 86.5% | 4.2% |
| Generalist (20 parts) | 80.4% | 84.5% | −4.1% (real > sim) |
| Full pipeline (perception) | N/A | 86.0% | N/A |
Limitations
- Author-stated: Dataset limited to 100 Autodesk assembly parts (primarily rigid, CAD-style geometries)
- Author-stated: Requires 3D-printable parts from the dataset; no flexible/deformable parts
- Unstated: Force-torque sensor and high-precision gripper hardware required; not broadly accessible
Reproducibility
- Code: Partial — dataset and simulation environments open-sourced
- Dataset: 100 assembly tasks from Autodesk, 3D-printable
- Compute: NVIDIA Isaac Sim on GPU cluster for RL training
Insights
The “real > sim” performance on the generalist policy (84.5% vs 80.4%) is an unusual result that suggests domain randomization overfit slightly to simulation artifacts. The real robot may benefit from consistent physics that the randomized simulator doesn’t perfectly replicate.
The 4.2% sim-to-real gap on precision assembly (gear insertion, tight-tolerance parts) is remarkably small and validates that domain randomization + imitation augmentation can bridge the gap even for contact-rich manipulation.
Connections
Raw Excerpt
“Assembly-by-Disassembly: instead of collecting 100 assembly demonstrations directly, we record 100 disassembly demonstrations and reverse them — bypassing the difficulty of manual assembly demonstration collection.”