Description
A robot arm opens doors, lids, and drawers it has never seen and identifies their mechanics (mass, friction, damping, stiffness) online, during the task, from its own sensing alone. What kinds of mechanics exist is learned offline; which mechanics this object has is estimated online by an exact Bayesian filter.
Project Details
A learned model predicts the object's motion as y = g(s)ᵀz + b(s), exactly linear in a per-object latent z. That makes a Kalman filter exact, so the robot gets a true posterior over the mechanics while the network stays frozen. The first columns of g are the physics regressor, so the latent stays physically interpretable.
What I built (MuJoCo, KUKA iiwa 14):
• Training on 13 simulated articulated objects with randomized mechanics, with the filter inside the training loss.
• A screw observer that finds the joint axis and reconstructs the joint's motion and force from the gripper alone, with no sensors on the object.
• A closed-loop open-and-close controller that turns with the estimated axis and adds short probing tugs only when the estimate is uncertain.
Results on never-seen objects:
• 650–3300× better 0.5 s motion prediction once the object is identified.
• Mass within 2× of truth in all 67 episodes (median 1.06–1.07×); friction median 0.99–1.03×.
• Every test object opens and closes in closed loop.
Presented as a poster at NERC. Next: using the estimate in the controller, and hardware. Supervised by Prof. Seth Hutchinson with Dr. Riddhiman Laha.