Humanoid robots have the potential for unparalleled versatility in performing human-like full-body skills. However, achieving agile and coordinated full-body movements remains a significant challenge due to the dynamic mismatch between simulation and the real world. Existing approaches, such as system identification (SysID) and domain randomization (DR) methods, often rely on laborious parameter tuning or result in overly conservative movements that sacrifice agility. Aligning Simulation and Real Physics (ASAP) may be a solution.

ASAP is a two-stage framework developed by NVIDIA in collaboration with Carnegie Mellon University that aims to bridge the gap between simulation and real physics when teaching agile full-body skills to humanoid robots.

In the first stage, the robot learns to move in a simulation using data acquired from human movements. This training takes place in a simulated environment with physical models that try to mimic reality as best as possible, but there are still differences between what the simulation predicts and how the robot behaves in reality.

In the second phase, the trained policies are transferred to the real world, where real data on the robot’s behavior is collected. Based on this, a so-called delta action model is created, which compensates for the differences between the simulated and real physics. This model is then used to fine-tune the robot’s movements, ensuring that its behavior in reality is as accurate and stable as possible.

In the video, you can see robots that imitate the movements of famous athletes.