OASIS: Observation-Action Space Alignment via SE(3) Trajectory Prediction for Robotic Manipulation

Researchers introduced OASIS, a visuomotor policy for robotic manipulation that fuses vision-language and depth features.

OASIS predicts camera-frame SE(3) end-effector trajectories, then uses pose-aware states to generate action chunks.

The method improves alignment between perception and control, helping robots reason about action-space geometry more directly.

In simulation and real-world tests, OASIS outperformed VLA and world action model baselines, with stronger success and generalization.