Carnegie Mellon University and NVIDIA have collaborated to create a new training method that enables humanoid robots to perform complex athletic movements with exceptional agility—such as Cristiano Ronaldo’s iconic mid-air spin celebration and Kobe Bryant’s fadeaway jump shot.

This innovative framework, called Aligning Simulation and Real Physics (ASAP), bridges the gap between simulated and real-world movements, allowing robots to carry out high-level athletic maneuvers that were once considered too complex for machines.

“Humanoid robots offer enormous potential for performing human-like, whole-body actions,” the researchers explained in their paper. “Yet, achieving coordinated and agile motions remains a major challenge due to discrepancies between simulation and the real world.”

ASAP addresses this challenge through a two-step process. First, it pre-trains motion tracking policies—algorithmic rules controlling the tracking—in a simulation using human motion data. Next, these policies are applied in real-world scenarios to gather data, helping to bridge the gap between simulated and real physics.

The outcome is a humanoid robot capable of replicating signature moves from sports legends, including Ronaldo’s “Siu” celebration (a 180-degree mid-air rotation), LeBron James’s “Silencer” celebration (involving precise single-leg balancing), and Kobe Bryant’s fadeaway jump shot (jumping and landing on one foot).

In addition to these notable sports moves, the robot also demonstrated impressive abilities like forward and side jumps exceeding 1 meter.

At first glance, the robots might seem somewhat clumsy, but this is largely due to hardware constraints, as they have much less articulation than humans. However, they are more dexterous than other robots, thanks to the “delta action model,” a correction mechanism that compensates for differences between simulated and real-world physics. “The delta action model serves as a residual correction term for the dynamics gap.”

By using this approach, the researchers reduced tracking errors by up to 52.7% compared to previous methods, enabling robots to perform complex movements once thought impossible.

“Our method significantly enhances agility and coordination across various dynamic motions,” the researchers noted, adding that their work is “laying the foundation for more versatile humanoid robots in real-world applications.”

Building robots with this level of dexterity has been one of the most challenging aspects of robotics. “For decades, we’ve imagined humanoid robots achieving, or even exceeding, human-level agility. However, most prior work has been focused primarily on locomotion, using the legs mainly for mobility,” the researchers wrote.

In contrast, ASAP mimics the human body during pre-training and then adapts this knowledge to real-world scenarios after being exposed to simulations. This allows the robot’s limbs to behave like human extremities, used for movement, balance, counterweight, and expression.

Achieving this is no small feat. When we perform athletic— or even basic— movements, we constantly make subtle adjustments, balancing multiple forces while compensating for momentum changes and shifting positions. Replicating this in robots has proven extraordinarily difficult.

If you doubt this, try playing the game QWOP, where you must control four body parts to make an athlete run. After mastering that, imagine managing the 21 basic articulations that ASAP handles, knowing that the human body has over 300 joints.

Humanoid robotics has been a growing field in recent years, with companies and universities dedicating more resources to research and development. Tesla’s Optimus project, Figure AI’s humanoid robot announcement, and Boston Dynamics’ Atlas have all shown increasing commercial interest in humanoid robots.

In academia, institutions like the University of Bristol and Stanford have developed their own methods to improve agility and dexterity in robots. The ASAP team plans to further develop the system.

“Future efforts could focus on damage-aware policy architectures to reduce hardware risks,” they said, referring to instances where models broke while performing complex moves. They also plan to research “leveraging markerless pose estimation or onboard sensor fusion to reduce reliance on MoCap systems” and refine adaptation techniques to increase efficiency. How long until we see an all-robot World Cup?