NVIDIA research shows robots trained in simulation can handle real-world tasks

NVIDIA Research announced eight accepted ICRA papers on sim-to-real robotics, showing simulation can improve real-world robot performance.

The studies cover navigation, grasping, flexible-object manipulation, assembly, clutter handling, and action-reasoning alignment.

Key methods include COMPASS, Grasp-MPC, Deformable Cluster Manipulation, SPARR, PEEK, and SEAL, built around NVIDIA Isaac Lab.

The results suggest simulation training is becoming a practical path to faster robot deployment in messy real-world settings.