Carnegie Mellon University and NVIDIA have partnered to create Aligning Simulation and Real Physics (ASAP). This AI framework allows humanoid robots to execute advanced athletic maneuvers, such as Cristiano Ronaldo’s iconic mid-air spin and Kobe Bryant’s signature fadeaway jump shot.

The two-stage framework uses pre-training motion tracking in simulation and refines policies with real data using a delta action model to align with real-world dynamics.

“Humanoid robots hold the potential for unparalleled versatility for performing human-like, whole-body skills. However, achieving agile and coordinated whole-body motions remains a significant challenge due to the dynamics mismatch between simulation and the real world,” said the team in the research paper.

Though researchers have previously attempted to develop robots with human-like agility, this has proved challenging due to hardware limitations and inconsistencies between real-world and simulated physics.

Three main approaches have emerged to combat this problem: System Identification (SysID), domain randomization (DR), and learned dynamics methods.

SysID, for example, estimates key physical parameters while relying on predefined spaces and torque measurements. DR trains policies in simulation with randomized parameters, often resulting in overly cautious movements. Though learned dynamics methods use real-world data to improve accuracy, the model’s effectiveness hasn’t been explored with humanoid robots.

The two-stage framework begins with pre-training policies in simulation using human motion videos, which are retargeted to humanoid robots and trained with a motion-tracking policy. The second stage collects real-world data to identify and correct discrepancies.

A “delta action model” is then trained to compensate for the detected differences between real-world and simulated physics. This correction mechanism reduced tracking errors by up to 52.7 percent.

This learning model has allowed ASAP-trained robots to replicate signature moves from sports legends, including Cristiano Ronaldo’s “Siu” celebration with a mid-air spin, LeBron James’s “Silencer” with precise single-leg balancing, and Kobe Bryant’s fadeaway jump shot. It can also perform front and side jumps exceeding one meter.

“Our approach significantly enhances agility and coordination across dynamic movements,” researchers noted in the study, highlighting its potential for real-world applications.