Pixel Aligned Robot Actions (PARA)
Are More Data-Efficient and Robust to New Viewpoints and Environments
Project Overview
Method Overview
PARA reformulated EEF coordinate regression as a per-pixel regression problem, enabling dramatically improved data efficiency and generalization compared to global coordinate regression, thanks to dense supervision and shift equivariance.
Results on Policy Trained on 40 Demonstrations (Same Viewpoint and Environment) (8x Speed)
97% Task Completion Rate
9% Task Completion Rate
5% Task Completion Rate
New Viewpoint Zero-Shot
52% Task Completion Rate
0% Task Completion Rate
0% Task Completion Rate
New Viewpoint with 5 Fine-tuning Episodes
87% Task Completion Rate
4% Task Completion Rate
4% Task Completion Rate
New Environment
94% Task Completion Rate
0% Task Completion Rate
6% Task Completion Rate
Other Tasks:
Wiping Table
95% Task Completion Rate
0% Task Completion Rate
61% Task Completion Rate
Fold Towel
97% Task Completion Rate
11% Task Completion Rate
42% Task Completion Rate