Humanoid robots are taking a giant leap forward thanks to a groundbreaking training technique developed by Carnegie Mellon University and NVIDIA. Their new framework, Aligning Simulation and Real Physics (ASAP), enables robots to execute high-level athletic maneuvers with agility never seen before. From Cristiano Ronaldo to Kobe Bryant, these robots are now replicating moves once thought impossible for machines.
Humanoid robots have long been envisioned as versatile machines capable of performing human-like full-body movements. However, achieving this level of agility has been a major hurdle due to the vast differences between simulated training environments and real-world physics.
The ASAP framework addresses this challenge through a two-stage approach. First, it pre-trains motion tracking policies in a simulated environment using real human motion data. Next, it deploys these policies in the physical world, collecting additional data to fine-tune performance and bridge the gap between theoretical and actual physics.
The result? A humanoid robot capable of nailing Cristiano Ronaldo’s iconic “Siu” celebration—a 180-degree mid-air rotation—alongside moves like LeBron James’s “Silencer” celebration, which requires precise single-leg balancing, and Kobe Bryant’s fadeaway jump shot.
Beyond sports-inspired maneuvers, these robots are demonstrating an impressive range of physical feats, including forward and side jumps over one meter. While they may still appear slightly clumsy due to hardware constraints, their level of dexterity is far ahead of previous robotic models.
A key innovation behind ASAP’s success is the “delta action model,” a correction mechanism that compensates for the inevitable differences between simulated and real-world physics. This model effectively serves as a real-time adjustment tool, reducing tracking errors by up to 52.7% compared to older methods.
“Our approach significantly improves agility and whole-body coordination across various dynamic motions,” the researchers stated, emphasizing that their system is a major step toward making humanoid robots more adaptable to real-world tasks.
Developing robots with human-like dexterity has been one of the toughest challenges in robotics. While previous research has largely focused on locomotion—using legs primarily for movement—ASAP takes a different approach, treating the entire body as an integrated system for motion, balance, and counterweight.
To put the difficulty into perspective, think about the viral game QWOP, where players struggle to control just four articulations to make an athlete run. Now, imagine managing the 21 basic articulations that ASAP handles—let alone the 300+ joints in the human body. This highlights just how complex robotic motion coordination truly is.
Interest in humanoid robots has skyrocketed, with both industry leaders and academic institutions investing heavily in research and development. Tesla’s Optimus project, Figure AI’s latest humanoid robot, and Boston Dynamics’ Atlas are all pushing the boundaries of what’s possible. Universities like Stanford and the University of Bristol are also exploring new ways to improve robotic agility and dexterity.
The researchers behind ASAP plan to refine the framework further. “Future directions could focus on developing damage-aware policy architectures to mitigate hardware risks,” they noted.
They also aim to reduce reliance on motion capture systems by leveraging markerless pose estimation and onboard sensor fusion.