Design a pick primitive with fully physics-based grasping for arbitrary objects in OmniGibson/BEHAVIOR-1K

Develop a pick action primitive that leverages fully physics-based grasping for arbitrary objects within the OmniGibson simulation used by BEHAVIOR-1K, enabling realistic grasp execution without the performance degradation observed when simplifying grasping during training and evaluation.

Background

In ablation experiments, enabling fully physics-based grasping at evaluation time caused a large drop in task performance, highlighting grasping as a critical bottleneck. To expedite training, the authors used an assistive pick primitive, but this diverges from realistic grasping dynamics.

The authors explicitly note that creating a pick primitive that supports fully physics-based grasping for arbitrary objects remains an open research problem, which is necessary to close the sim-real gap and to evaluate policies under realistic manipulation dynamics.