In the realm of robotics, achieving true agility has been akin to chasing a mirage—until Cassie came along. While robots performing parkour in viral videos might suggest we’ve cracked the code, the reality has been far less adaptable. Enter a groundbreaking approach that could finally make robots nimble enough to handle the unexpected: reinforcement learning.

Teaching Old Bots New Tricks

Cassie, a two-legged robot, has shattered expectations by mastering the art of running 400 meters across varied terrains, along with performing impressive standing long jumps and high jumps. This wasn’t achieved through traditional programming but by leveraging reinforcement learning, a method that rewards or penalizes the robot as it attempts different objectives. This technique enabled Cassie to generalize its responses to new scenarios, marking a departure from the limitations of its predecessors.

Fast-Track to Flexibility

The secret sauce in Cassie’s training was a simulation environment that accelerated its learning process from years to mere weeks. Starting with basic movements like jumping and running, Cassie was initially encouraged to mimic human movements and animations. As it mastered each task, the training evolved through task randomization, preparing Cassie for unforeseen challenges—like maintaining balance while being pulled sideways.

The Future on Two Legs

Cassie’s achievements, including a 400-meter dash in just two minutes and 34 seconds and a 1.4-meter long jump, are just the beginning. The team is now exploring how to integrate on-board cameras for even more complex interactions with the world. The vision? Humanoid robots that can not only move with grace but also perform real work, interact dynamically with their environment, and possibly revolutionize the field of robotics. Cassie’s journey from simple steps to dynamic motions underlines a pivotal shift in robot agility, driven by AI’s endless possibilities.

SOURCE: How AI taught Cassie the two-legged robot to run and jump | MIT Technology Review