Coaching Robots For Athletic Actions

Can robots transfer like athletes? A brand new coaching mannequin helps them replicate sports activities strikes, however the outcomes present each progress and surprising challenges.

A staff of AI and robotics researchers from Carnegie Mellon College, together with two colleagues from NVIDIA, has created a brand new mannequin to coach robots to maneuver like human athletes. The staff noticed that the majority robotic coaching focuses on locomotion, resulting in robots that transfer effectively however with out fluidity or athleticism. To deal with this, they explored whole-body coaching. They discovered present fashions lacked adaptability and relied on too many parameters, making robotic actions overly cautious. This led them to develop a brand new two-stage coaching framework.

The primary stage trains an AI module to research whole-body human movement movies, adjusting key actions to suit the robotic’s capabilities utilizing movement monitoring. The second stage gathers real-world information to bridge the hole between human motion in movies and the way robots can bodily transfer. This course of led to a framework known as Aligning Simulation and Actual Physics (ASAP).

The ASAP framework consists of 4 steps. First, movement monitoring pre-training and actual trajectory assortment contain retargeting humanoid motions from human movies. A number of movement monitoring insurance policies are pre-trained to generate real-world motion trajectories. Subsequent, delta motion mannequin coaching is carried out utilizing real-world rollout information. This step minimizes the discrepancy between the simulated state and the precise real-world state, enhancing the mannequin’s accuracy.

Within the coverage fine-tuning stage, the delta motion mannequin is frozen and built-in into the simulator to higher align with real-world physics. The pre-trained movement monitoring coverage is then fine-tuned for larger precision. Lastly, in real-world deployment, the fine-tuned coverage is applied straight in the true world with out counting on the delta motion mannequin, making certain the robotic can carry out its skilled actions independently.

To check the framework, the researchers skilled a robotic to duplicate iconic sports activities strikes. It carried out Kobe Bryant’s fadeaway bounce shot, LeBron James’ Silencer transfer, and Cristiano Ronaldo’s Siu leap with a mid-air spin. Every motion was recorded.

The robotic’s actions clearly resemble the well-known sports activities strikes, highlighting progress in full-body movement. Nonetheless, it’s additionally evident that rather more work is required earlier than a robotic could possibly be mistaken for an expert athlete.

Reference: Tairan He et al, ASAP: Aligning Simulation and Actual-World Physics for Studying Agile Humanoid Complete-Physique Abilities, arXiv (2025). DOI: 10.48550/arxiv.2502.01143