Boston Dynamics recently announced that its quadruped robot, Spot, has mastered the challenging maneuver of performing consecutive backflips through reinforcement learning. According to technology media outlet Notebookcheck, while this skill wasn't directly requested by customers, the training behind it significantly improved the robot's survivability in extreme situations like falls and slips, allowing it to effectively adjust its posture to protect its body and the expensive sensors on its back.
The development process was fraught with challenges. The team first simulated the backflip in a computer, but engineer Arun Kumar revealed that initial attempts to apply the simulation results to the real robot failed almost every time. To mitigate risk, testing began on a gymnastics mat and gradually increased the difficulty, ultimately enabling Spot to reliably perform the maneuver in this high-risk environment. Surprisingly, after mastering the backflip, Spot's walking gait became more natural, resembling the movement patterns of a real quadruped.
Reinforcement learning was at the heart of this breakthrough. Through extensive trial and error and feedback-based adjustments, Spot not only mastered precise body control but was even able to maintain balance while equipped with rollers on its front legs. This advancement demonstrates the enormous potential of AI in robotic control and lays the foundation for future applications in complex environments.
