NVIDIA Research Extends Robotics Technology from Simulation to the Real World

Robotics is moving beyond automation that repeats fixed motions, entering the stage of Embodied Autonomy, where robots judge and move on their own even in unpredictable real-world environments. At ICRA, NVIDIA Research presented a range of research that transfers simulation technology to the real world (Sim-to-real) to enhance robots' perception, reasoning, and planning capabilities.

Key research achievements are as follows:

• ScheduleStream: Uses GPUs to plan the movements of multiple robot arms in parallel. It achieves 3x faster speed compared to existing methods and can run on the NVIDIA Jetson platform.
• COMPASS: A framework that enables navigation even when robot shapes differ. It learns through reinforcement learning within NVIDIA Isaac Lab, boosting the success rate by 4.5x compared to existing methods, and recorded a success rate of about 80% in real-world tests.
• Grasp-MPC: An adaptive grasping technology that corrects movements in real time when approaching objects. Leveraging the cuRobo library and 2 million simulated trajectories, it achieved a success rate of 75%, far surpassing existing methods even in complex environments.