Humanoids that have demonstrated this capability before typically separate the low-level walking and balancing control from the control of the arms for manipulation. In the TRI demonstration, a single LBM had direct control of the entire robot, treating the hands and the feet almost identically.

“One of the main value propositions of humanoids is that they can achieve a huge variety of tasks directly in existing environments, but the previous approaches to programming these tasks simply could not scale to meet this challenge,” says Russ Tedrake, senior vice president of large behavior models at TRI.

“Large behavior models address this opportunity in a fundamentally new way,” explains Tedrake. “Skills are added quickly via demonstrations from humans, and as the LBMs get stronger, they require less and less demonstrations to achieve more and more robust behaviors.”

According to Tedrake, by adopting LBMs, new capabilities that previously would have been laboriously hand-programmed can now be added quickly and without writing a single new line of code.

“This work provides a glimpse into how we’re thinking about building general-purpose robots that will transform how we live and work,” adds Scott Kuindersma, vice president of robotics research at Boston Dynamics. “Training a single neural network to perform many long-horizon manipulation tasks will lead to better generalization, and highly capable robots like Atlas present the fewest barriers to data collection for tasks requiring whole-body precision, dexterity and strength.”