Lexi Foland, Thomas Cohn, Adam Wei, Nicholas Pfaff, Boyuan Chen, and Russ Tedrake
Diffusion policies have shown impressive results in robot imitation learning, even for tasks that require satisfaction of kinematic equality constraints. However, task performance alone is not a reliable indicator of the policy's ability to precisely learn constraints in the training data. To investigate, we analyze how well diffusion policies discover these manifolds with a case study on a bimanual pick-and-place task that encourages fulfillment of a kinematic constraint for success. We study how three factors affect trained policies: dataset size, dataset quality, and manifold curvature.
Software Engineering Internship
Over the course of 10 weeks, I worked alongside engineers at IHMC to create a teleoperation system for two robotic arms. The teleoperation system relied on markerless motion capture software, in which cameras identify a person's arm and shoulder linkages. Our system sent commands to the robotic arms that mimicked these real-world poses.
Links: LinkedIn Post
Final Project, 6.4210 (Robotic Manipulation) at MIT
We propose a system that involves three robots: two robotic arms serve as “designer” arms that phys- ically manipulate a third, passive “mannequin” humanoid. We use compliant control to enable the contact-rich interactions required to pose the humanoid in desired configurations. We leverage strategies such as antipodal grasp computation, updating trajectory planning, and control with gravity compensation and velocity damping to improve task success rates. Finally, we present a series of nontrivial target configurations that our system can mimic via posing of the passive humanoid with the robotic arm and an analysis of the system’s ability to mimic arbitrary configurations.
Links: YouTube Video