Robot mimics athlete iconic sports actions with new whole-body training model

A team of AI and robotics researchers at Carnegie Mellon University, working with a pair of colleagues from technology company NVIDIA, has developed a new model for training robots to move like human athletes.

In their paper posted on the arXiv preprint server, the group describes how they developed the new approach to allow for training full-body athletic movements with humanoid robots, and how well the approach has worked thus far.

In their new effort, the research team noted that most efforts to train robots to do things center mainly around locomotion. The result has been the development of a host of robots that are able to get around very well. But none of them, the team notes, do it with much grace; they lack fluidity or athleticism—hallmarks of natural animal movements. The answer, they believed, was to shift the focus to using whole-body training.

In looking to develop whole-body training, the team found that current training models lacked adaptability and often used too many parameters, resulting in overly cautious movements. That led them to develop a new two-stage model, or framework as they call it. 

The first stage involves training an AI module to understand whole-body human motion videos—with the salient points retargeted to consider robot capabilities in conjunction with motion tracking. The second stage involves collecting real-world data to identify and reconcile differences between actions in the real world (the way people move in the videos) and how robots can move. The result is a framework the team calls Aligning Simulation and Real Physics (ASAP).

To test the new framework, the researchers trained a robot to make moves familiar to sports fans. The robot performed Kobe Bryant's famous fadeaway jump shot, LeBron James' Silencer move and Cristiano Ronaldo's Siu leap with a mid-air spin. Each whole-body skill was recorded as it was performed, and the results were posted to YouTube.

Watching them, it is easy to recognize the famous moves and note the progress made in improving full-body motion. But it is also easy to see that much more work needs to be done before a robot will ever be mistaken for a professional human athlete.

Retargeting Human Video Motions to Robot Motions: (a) Human motions are captured from video. (b) Using TRAM [93], 3D human motion is reconstructed in the SMPL parameter format. (c) A reinforcement learning (RL) policy is trained in simulation to track the SMPL motion. (d) The learned SMPL motion is retargeted to the Unitree G1 humanoid robot in simulation. (e) The trained RL policy is deployed on the real robot, executing the final motion in the physical world. This pipeline ensures the retargeted motions remain physically feasible and suitable for real-world deployment. Credit: arXiv (2025). DOI: 10.48550/arxiv.2502.01143

by Bob Yirka , Tech Xplore

Source: Ronaldo's Siuuu celebration: Whole-body training model allows robots to mimic famous athlete moves