Pixel Aligned Robot Actions (PARA)
Are More Data-Efficient and Robust to New Viewpoints and Environments
Project Overview
Method Overview
We reformulate robot actions as a per-pixel regression problem, enabling dramatically improved data efficiency and generalization compared to global coordinate regression, thanks to dense supervision and shift equivarianc
Data Efficiency
Same Viewpoint and Environment; Ours vs ACT.
Pick and Place
97% Task Completion Rate
9% Task Completion Rate
—% Task Completion Rate
—% Task Completion Rate
Fold Towel
97% Task Completion Rate
11% Task Completion Rate
—% Task Completion Rate
—% Task Completion Rate
Wipe Table
95% Task Completion Rate
0% Task Completion Rate
—% Task Completion Rate
—% Task Completion Rate
Robustness to Viewpoint and Environment
52% Task Completion Rate
0% Task Completion Rate
87% Task Completion Rate
4% Task Completion Rate
94% Task Completion Rate
0% Task Completion Rate
—% Task Completion Rate
—% Task Completion Rate
—% Task Completion Rate
—% Task Completion Rate
—% Task Completion Rate
—% Task Completion Rate
Deployed on Other Embodiments
SO300
Piper
Panda
G1
Scaling up to Calibrated Droid
Put Toy into Box
PARA (Ours) — % Completion Rate
ACT — % Completion Rate
Press Button
PARA (Ours) — % Completion Rate
ACT — % Completion Rate
Pick Up Sponge and Place in Bowl
PARA (Ours) — % Completion Rate
ACT — % Completion Rate
More Data-Efficient Hand Co-Training
Hand Demonstrations Curtained with PARA (Ours) vs ACT.
50% Task Completion Rate
80% Task Completion Rate
80% Task Completion Rate
100% Task Completion Rate
5% Task Completion Rate
10% Task Completion Rate
15% Task Completion Rate
20% Task Completion Rate
Backbones
DINO
95% Task Completion Rate
SmolVLA
87% Task Completion Rate
MoGe
92% Task Completion Rate
COSMOS Video Model
94% Task Completion Rate