Researchers Enable Humanoid Robot to Imitate Iconic Athlete Moves

A collaborative research team from Carnegie Mellon University and Nvidia has unveiled an innovative AI framework called Aligning Simulation and Real Physics (ASAP), designed to enhance the agility of humanoid robots. This framework allows the Unitree G1 robot to replicate the signature moves of renowned athletes such as Cristiano Ronaldo and LeBron James, pushing the boundaries of robotic motion.

Humanoid robots have historically struggled with executing dynamic and coordinated whole-body movements. Most agile maneuvers have been limited to basic locomotion or simple tasks, with past attempts to implement full-body agility hampered by hardware limitations and discrepancies between simulated and real-world physics. To address these challenges, the research team developed the two-tiered ASAP system, which employs video data of human movements for initial training in a simulated environment.

Once the robot has been trained in the simulation, the strategies are transferred to the humanoid robot, which is then subjected to a motion-tracking regimen that aligns with real-world data. This approach involves the training of a delta action model to reconcile the differences between simulated training and actual execution, enabling the robot to adapt to real physical conditions. Consequently, the robot exhibits improved agility in its movements.

In their experiments, the researchers utilized the Unitree G1 robot to test the ASAP framework, teaching it to perform moves inspired by famous athletes. This included Cristiano Ronaldo's celebrated 'Siu' jump, the one-legged balancing technique of LeBron James known as the 'Silencer', and Kobe Bryant's signature 'Fadeaway Jumper'. The researchers also trained the robot to execute various athletic maneuvers, including forward and lateral jumps of up to one meter.

Through the implementation of the ASAP framework, the researchers noted significant improvements in the robot's ability to perform complex whole-body movements. While the robot's actions still appear somewhat clumsy, this is attributed to the hardware constraints of the Unitree G1, which has fewer degrees of freedom in its joints compared to human anatomy.

The advancements made through the ASAP framework signal a substantial step forward in robotic agility and movement, potentially paving the way for more sophisticated humanoid robots in the future.