Parkour

A legged robot's high-dynamic skill of continuously climbing, jumping, and squeezing through obstacles using vision.

Parkour was originally a human street sport; in robotics it refers to a quadruped or humanoid robot rapidly crossing platforms, gaps, low barriers, and narrow openings, and serves as a flagship task for testing the limits of motion control. Boston Dynamics' hydraulic Atlas parkour demos used motion templates built offline through trajectory optimization, fine-tuned in real time with model predictive control and perception. Since 2023, learning-based methods have become the mainstream: Robot Parkour Learning (CoRL 2023) and Extreme Parkour first train with reinforcement learning in simulation, then distill the result into an end-to-end policy that uses only a depth camera, deployed directly on low-cost quadrupeds; Humanoid Parkour Learning (2024) applied a similar approach to humanoid robots.

ExampleIn Extreme Parkour, a low-cost quadruped, relying only on a front-facing depth camera, jumps onto a platform about twice its body height and clears a gap about twice its body length.

Also called

Robot Parkour

Related

Legged Locomotion · Perceptive Locomotion · Robot Parkour Learning · Extreme Parkour · Sim-to-Real Transfer · Teacher-Student Distillation

Sources

Robot Parkour Learning
Extreme Parkour with Legged Robots
Flipping the Script with Atlas (Boston Dynamics)

As of

2024-06