Practical Applications of Embodied AI and Humanoid Robots

Engineering & simulationDiffusion Policy / Fusion Policy (TRI)100 demonstrations

TRI researchers used a diffusion policy derived from image generation techniques to decode robot actions, allowing the robot to learn complex dexterous tasks like folding shirts or beating eggs after approximately 100 teleoperation demonstrations.

“we had a result with Sharon song in collaboration with Sharon song. It was an internship project that started here, and then we invested heavily once it started looking very promising, called the fusion policy, which was just taking the state of the art ideas from image generation and trying to transition that to producing robot actions.”

Engineering & simulationBoston Dynamics Machine Learning Team

Boston Dynamics developed a repeatable process to turn human demonstrations of manipulation tasks into autonomous skills using diffusion policy techniques, aiming to improve robustness and generalizability in real-world environments.

“we found that that's really powerful, right? So basically, uh, we now are in a place where we can demonstrate, um, manipulation tasks on a robot. We have a pretty repeatable process for turning those into autonomous skills, and that's a really powerful thing.”