Boston Dynamics Trains Atlas Robot to Perform Complex Tasks on Command in Real Time
Credit: Boston Dynamics
Boston Dynamics has unveiled an impressive leap forward in humanoid robotics, showing how artificial intelligence enables the Atlas robot to not only move, but also perform long, complex manipulations of objects—all under the control of learnable large behavioral models (LBMs) created in partnership with Toyota Research Institute (TRI).
In a video released recently, Atlas performs complex tasks without a single line of code – all of its movements are based solely on learning.
In particular, in the video, the robot is engaged in packing, sorting and placing objects, combining hand manipulation and full-body movement. The integration of large behavioral models, which became the basis of the work of Boston Dynamics and TRI, allows adding new skills without writing code, whereas previously this required lengthy programming. This breakthrough was made possible by combining the expertise of Boston Dynamics in robotics and TRI in the field of artificial intelligence. This symbiosis opened the way to the creation of universal humanoid assistants. This is especially important since humans have adapted the world to themselves – this means that a robot will behave in it most naturally if it resembles a person and can repeat his movements and gestures.
According to the project managers, the main innovation is that large behavioral models allow for a large-scale expansion of the range of tasks and the rapid addition of new skills, making robot behavior truly universal. The company continues to experiment, increase the volume and complexity of tasks, and introduce new algorithmic approaches to make robots even more reliable and autonomous.
