AutoMate: Specialist and Generalist Assembly Policies over Diverse Geometries
July 13, 2026
Training Specialist Policies:
1.Assembly-by-Disassembly
-collect demonstration for disassembly instead and reverse the disassembly paths for assembly
2.Reinforcement Learning w/ Imitation Objective
-use PPO with engineered reward and imitation reward :
- based on the closest demonstration path
3.Dynamic Time Warping (DTW) for Trajectory Matching
-find a mapping between end effector path and demonstration path to fix discretization issues between sampling rates
Recipe:
1.Standard behavior cloning baseline
2.RL fine-tuning for assembly tasks