Oh, what a feeling: Toyota building robots that get better with practice

Bots that learn to peel potatoes is a lot less scary than Black Mirror

Boston Dynamics and Toyota Research Institute (TRI) announced on Wednesday they're partnering to combine the former's multi-jointed athletic humanoid, Atlas, with TRI's large behavior models (LBM).

Boston Dynamics CEO Robert Playter enthused that he was looking forward to accelerating "the development of general-purpose humanoids," while TRI CEO Gill Pratt cheered that "recent advances in AI and machine learning hold tremendous potential for advancing physical intelligence."

As a reminder, this is Atlas:

Youtube Video

TRI's LBM work includes the generative AI technique known as diffusion policy, which allows robots to acquire new dexterous behaviors by having them demonstrated, rather than programmed. What that means, according to one video, [VIDEO] is that robots can learn to do tricky things like peel potatoes and flip pancakes.

Diffusion policies help a robot learn how to perform fine motor skills by generating small, sequential actions that gradually build up to a more complex behavior. Instead of predicting a single, definitive action in one step, the diffusion policy predicts an array of possible actions and gradually narrows them down over time – allowing the robot to become more accurate at handling objects.

The LBMs are based on combining the skills created via the diffusion policy efforts.

"This allows us to teach robots skills faster and with significantly fewer demonstrations," explained TRI VP Russ Tedrake.

Before diffusion policies, most robotic object manipulation focused on "pick and place tasks," which limited robots to simple objects and rearranging, commented TRI's Ben Burchfiel. Diffusion policies can make the most of hardware capabilities such as touch sensors without modifying any code or explicitly programming any new skills, he added.

One action can be taught to a machine by a human in the afternoon, then the robot is left alone to practice the action and by the next morning, the machine has learned the action. Importantly, once one robot has mastered a skill, that knowledge can be deployed to a fleet of robots at once.

TRI has referred to its training as a "kindergarten for robots."

"It's just amazing to see the tasks that the robots can perform. Even a year ago, I never would have expected that robots would become this skilled," remarked Tedrake.

Tedrake co-leads the partnership along with Boston Dyanmics' senior director Scott Kuindersma.

"The physical capabilities of the new electric Atlas robot, coupled with the ability to programmatically command and teleoperate a broad range of whole-body bimanual manipulation behaviors, will allow research teams to deploy the robot across a range of tasks and collect data on its performance," stated TRI.

That data, said the company, will go back into training more advanced LBMs. The joint team's future research will then focus on answering "fundamental training questions for humanoid robots, the ability of research models to leverage whole-body sensing, and understanding human-robot interaction and safety/assurance cases to support these new capabilities."

Here's hoping for a focus on the safety part, as humanity is currently in an era where similar AI-empowered robots have been seen toting machine guns.

There have been expressions of concern from researchers on efforts to equip robots with AI – including this February when computer scientists at the University of Maryland (UMD) warned that "it is easy to manipulate or misguide the robot's actions, leading to safety hazards." ®