Empowering Robots for Swift, Smart Decision-Making

In 2018, Google DeepMind’s AlphaZero program demonstrated its ability to master chess, shogi, and Go through self-directed learning and a specialized algorithm that assessed optimal moves on a defined grid. Now, researchers at Caltech have adapted this idea for autonomous robots, crafting a planning and decision-making control system that enables these robots to navigate their surroundings effectively.

Soon-Jo Chung, Caltech’s Bren Professor of Control and Dynamical Systems and a senior research scientist at JPL (managed by Caltech for NASA), notes, "Our algorithm strategizes by exploring various significant motions, ultimately selecting the best one through dynamic simulation." Their innovative approach is termed Spectral Expansion Tree Search (SETS), as detailed in the December edition of Science Robotics.

Robots, like humanoid assistants for the elderly, must maneuver freely, adapting to obstacles and unexpected situations. John Lathrop, a graduate student and co-lead author, emphasizes the need for SETS to automate decision-making rather than relying on designers to dictate movements. The algorithm employs control theory and linear algebra to maximize a robot’s physical capabilities.

SETS employs a Monte Carlo Tree Search strategy, where potential moves are represented as nodes branching from a central idea. However, in continuous systems, the number of trajectories can quickly become overwhelming. To address this, SETS strikes a balance between exploring new routes and exploiting known effective ones.

Running in just a tenth of a second, SETS can simulate thousands of trajectories to identify the optimal path. It functions across various robotic applications, evidenced by its success in diverse experiments—from drone navigation to assisting drivers and coordinating tethered spacecraft. This innovative work, supported by esteemed agencies, paves the way for future advancements in robotics.