Robots can now train themselves with new "practice makes perfect" algorithm

MIT researchers develop EES algorithm to help robots autonomously improve skills through practice, tested on Boston Dynamics' Spot for efficient learning.

: MIT's CSAIL and The AI Institute presented the 'Estimate, Extrapolate, and Situate' (EES) algorithm at the Robotics: Science and Systems conference. The algorithm allows robots to autonomously identify and practice improving their weaknesses. Tested on Boston Dynamics' Spot, EES notably reduced the time needed for skill mastery. Researchers aim to integrate simulators for combined virtual and physical practice sessions in the future.

MIT's Computer Science and Artificial Intelligence Lab (CSAIL) and The AI Institute have created the 'Estimate, Extrapolate, and Situate' (EES) algorithm, recently showcased at the Robotics: Science and Systems conference. This innovative approach enables robots to autonomously assess their skills, determine areas needing improvement, and systematically practice to enhance their capabilities.

Researchers tested EES on Boston Dynamics' Spot robot, known for its proficiency in various tasks, especially when equipped with an arm attachment. The EES algorithm significantly reduced the time required for Spot to master new skills, such as placing a ball on a slanted table in three hours and sweeping toys into a bin in two hours, compared to over 10 hours with previous methods.

Looking ahead, the researchers plan to integrate simulators to combine virtual and physical practice sessions for faster learning. They also aim to develop algorithms that can reason over sequences of practice attempts, aiding robots in performing a wide range of tasks autonomously. This advancement is seen as crucial for future home robots expected to learn on the job, as emphasized by Danfei Xu from Georgia Tech and Nvidia AI.