Russ Tedrake (MIT): "Learning Manipulation — and Why I (Still) like F=ma" - 2019

Details

Title : Russ Tedrake (MIT): "Learning Manipulation — and Why I (Still) like F=ma" Author(s): MIT Institute for Data, Systems, and Society Link(s) : https://www.youtube.com/watch?v=Tyz1jRfyWgY

Rough Notes

Speaker was originally working in Reinforcement learning (RL), but now more focused on (robotic) manipulation. The demo shows a robot that puts things into a dishwasher, which is harder than just picking up an object and putting it somewhere, since we would need to put it to a specific place in the dishwasher.

One problem with manipulation problems is that they break the rigorous/reliable approaches (that the speaker knows) for control, albeit in important and exciting ways. Nobody uses principled feedback control in manipulation as well. When we find a failure case, it has to be fixed manually.

Non-smooth mechanics for contact has stiff differential equations with discontinuous/stiff contact forces which are set-valued (e.g. due to Coulomb friction). If we a have a plan of what the robot and the plate are gonna do, how do we robustly stabilize it.

The Model Predictive Control (MPC) method for contact mechanics fails as linearization cannot capture even the local (non-smooth) dynamics. A locally valid approximation looks like a piecewise-affine (PWA) system. This is formulated as a mixed-integer convex optimization problem. This is however too computationally intensive. Kye ideas for the tight formulations for PWA MPC is to have a convex hull formulation for subgroups of decision variables, use the objective function in the convex hull (see Mixed-Integer Formulations for Optimal Control of PWA Systems, Marcucci and Tedrake for more information).

Emacs 29.4 (Org mode 9.6.15)