Humanoid Robots Take a Leap Forward with New Training Technique
Carnegie Mellon University and NVIDIA have unveiled a groundbreaking training method that enables humanoid robots to execute highly agile athletic maneuvers, pushing the boundaries of what machines can physically achieve. From mimicking Cristiano Ronaldo’s mid-air spin celebration to executing Kobe Bryant’s fadeaway jump shot, robots are now capable of performing movements that were once thought impossible.
This advancement is made possible through a framework known as Aligning Simulation and Real Physics (ASAP), which bridges the long-standing gap between simulated motion training and real-world execution. Historically, humanoid robots have struggled with high-level agility and coordination due to the inherent differences between digital models and actual physics. ASAP addresses this issue by employing a two-stage learning process. First, the robots are pre-trained using motion data from human movements in a simulated environment. Then, they refine these learned motions in the real world, using newly collected data to improve precision and coordination.
The results are remarkable. Robots can now perform intricate sports moves with an unprecedented level of control, including Ronaldo’s famous "Siu" celebration, LeBron James’ "Silencer" pose, and even high jumps exceeding one meter. While they may still appear somewhat clumsy due to hardware limitations, these machines are significantly more agile than their predecessors, thanks to a "delta action model" that compensates for discrepancies between simulation and reality. This innovation has reduced motion tracking errors by over 50%, allowing for smoother and more natural movements.
The research team believes this advancement brings humanoid robots one step closer to real-world applications beyond research labs. Unlike previous robotic models, which primarily focused on simple locomotion, ASAP enables robots to utilize their entire bodies for balance, momentum control, and even expressive movements. The challenge lies in replicating the countless micro-adjustments that humans naturally make while moving—something that even the most advanced robots have struggled with for decades.
The field of humanoid robotics has gained significant traction in recent years, with companies like Tesla, Figure AI, and Boston Dynamics pushing the envelope on robotic development. Universities worldwide, including Stanford and the University of Bristol, have also invested in refining robotic agility and dexterity.
Looking ahead, researchers aim to enhance ASAP further by integrating damage-aware policies to protect hardware from excessive strain. They are also exploring new ways to improve adaptation efficiency, reduce reliance on motion capture systems, and fine-tune their models for more advanced applications.
With each breakthrough, humanoid robots edge closer to matching human agility. If progress continues at this pace, we may one day see a fully robotic sports league—or even an all-robot World Cup.
