NVIDIA Research Advances Robotics From Simulation to the Real World
NVIDIA Blog · Katie Washabaugh · 2026-05-28
NVIDIA Research presents eight papers at ICRA 2026 demonstrating how simulation-to-real transfer enables robots to generalize across multi-arm coordination, novel object grasping, precise assembly, and vision-language-guided action execution with substantial real-world performance gains.
Appears in
Extraction
Topics: roboticssim-to-realrobot-learningphysical-aivision-language-action
Claims
- ScheduleStream uses GPU-parallelized computation to achieve a 3x speedup in multi-arm robot planning scenarios on hardware including NVIDIA Jetson.
- COMPASS trains navigation policies entirely in simulation via imitation and residual reinforcement learning, achieving approximately 80% success across 20 real-world trials on diverse robot embodiments.
- Grasp-MPC continuously corrects grasping trajectories in real time and achieves approximately 75% success on novel objects in clutter versus a 41% baseline.
- SPARR separates simulation-trained strategy from hardware-specific correction to improve robot assembly success by 38% and reduce cycle time by 30% compared to zero-shot sim-to-real baselines.
- The PEEK pipeline uses a vision-language model to focus robot attention on task-relevant objects, achieving a 41x real-world accuracy improvement for policies trained purely in simulation.
Key quotes
Sim-to-real is becoming a foundation for robots that can adapt, generalize, and operate with greater reliability outside the lab.
SEAL fixes this at runtime without any retraining: the robot generates several candidate action sequences, thinks through where each one would actually lead and picks the outcome that matches what it said it would do.
The NVIDIA Physical AI Dataset is the world's largest open dataset for physical development, surpassing 15 million+ downloads.