Researchers have developed a new framework that allows humanoid robots to perform parkour-style movements, bringing them closer to human levels of agility in complex environments, tells Tech Xplore. The system, called Perceptive Humanoid Parkour (PHP), enables robots to run, jump, climb, and traverse obstacles while dynamically adapting to changes in their surroundings. The work was created by a research team from Amazon Frontier AI & Robotics and the University of California, Berkeley.
Humanoid robots, designed with human-like bodies and joints, are considered promising tools for tasks that currently require people, especially in unpredictable environments such as disaster zones, warehouses, or construction sites. However, achieving the flexibility and coordination needed for such tasks has been a persistent challenge in robotics. The new PHP framework addresses this problem by combining perception, learning, and motion planning in a single system.
The approach relies on a multi-skill visuomotor policy that allows the robot to select and transition between different athletic movements. Through a training process that combines motion matching and a teacher–student learning pipeline, the robot learns to replicate a variety of dynamic actions. These include cat-vaulting over obstacles, speed-vaulting while running at roughly three meters per second, and climbing walls as tall as 1.25 meters, almost the height of the robot itself.
Researchers tested the system on a Unitree G1 humanoid robot. In experiments, the machine successfully navigated a complex parkour course for about 60 seconds, autonomously choosing appropriate skills and transitioning smoothly between them. The robot could also adjust its movements in real time when obstacles shifted or unexpected conditions occurred.
The study highlights the importance of integrating perception with control in robotics. By enabling robots to understand their surroundings and adapt their movements instantly, the PHP framework represents a step toward more capable humanoid machines.
Such advances could eventually allow humanoid robots to operate in cluttered real-world environments where agility, balance, and rapid decision-making are essential.