Boston Dynamics teams up with former CEO's research institute to use reinforcement learning to improve Atlas, a humanoid robot

Boston Dynamics, known for developing the bipedal humanoid robot 'Atlas,' has announced a partnership with the nonprofit Robotics & AI Institute (RAI Institute) to further evolve 'Atlas' through reinforcement learning. The RAI Institute was founded in 2022 by Mark Raibert, who served as CEO of Boston Dynamics for 30 years.

Boston Dynamics and the RAI Institute Partner | Boston Dynamics

Boston Dynamics joins forces with its former CEO to speed the learning of its Atlas humanoid robot | TechCrunch
https://techcrunch.com/2025/02/05/boston-dynamics-joins-forces-with-its-former-ceo-to-speed-the-learning-of-its-atlas-humanoid-robot/

According to the announcement, Boston Dynamics will collaborate with the RAI Institute to establish a shared reinforcement learning training pipeline for the new Atlas and build dynamic, generalizable mobile manipulation behaviors, starting in February 2025.

Boston Dynamics and RAI Institute have previously collaborated on the development of a reinforcement learning kit for the quadruped robot Spot. The kit for Spot has successfully run at 11.5 miles per hour.

This project, which focuses on 'Atlas,' has three objectives:

1: Development of 'Sim to Real' for robots
Despite the existence of fast parallel simulators and sophisticated control optimization techniques, 'applying simulation results to real hardware' remains one of the most challenging aspects of robotics. To bridge this 'simulation-reality gap', the two teams are working together to train policies that perform a variety of agile movements on physical hardware, aiming to achieve novel, robust, and practical locomotion (movement) movements.

2: Improve whole body locomotion
The robot's usefulness would increase if it had the ability to manipulate objects such as doors or levers in conjunction with locomotion, and the team plans to develop policies to improve stability in these cases.

3. Explore whole-body exercise strategies
Third, we will explore tasks that require full-body movement strategies that require tight coordination between the arms and legs, such as running fast or handling heavy objects, and advanced full-body locomotion. The goal is to use reinforcement learning to generate such behaviors during complex contact events without imposing strict requirements.

'We live in an exciting time for humanoid robotics development. But to be useful, humanoid robots need to be flexible enough to operate in a variety of environments and perform tasks across a wide range of applications,' said Robert Player, CEO of Boston Dynamics. 'Our partnership with the RAI Institute brings together two of the world's leading robotics organizations to help strengthen the core capabilities needed to make robots like Atlas a valuable part of people's lives.'