Videos

RieMO running as the core optimization process of a continuously optimizing control system for Apollo demonstrating movement agility and robustness to perturbations

All motions are generated on the fly by optimization (RieMO) running in real time. Uses force feedback during grasp measured by the Barrett hand's strain gauges.

Riemannian Motion Optimization (RieMO): Simulation demonstrations of optimized motions around, and in contact with, obstacles under different workspace metrics we've studied. Each of these motions took around .7 seconds to optimize from a naïve zero motion trajectory using code designed for usability and correctness over speed.

Motion Demos: Basic motion demonstrations on the physical Apollo platform. Each motion segment between zero velocity states is optimized using the RieMO framework. The motions use various combinations of approximate energy measures, workspace and configuration space velocity penalties, bias configuration redundancy resolution penalties, workspace motion biasing potentials (e.g. grasp and lift potentials for manipulation), joint limit penalties and constraints, environmental constraints (obs avoidance), end-effector orientation constraints, and motion target constraints.

Dual Execution: Handling modeling uncertainty formally in motion generation is computationally intractable. However, empirically we've seen that contact controllers can be surprisingly robust to uncertainty and perturbations on their own. This framework adds potentials to the motion optimizer to design motion that explicitly reduces uncertainty through contacts. Information about those desired contacts is communicated to lower level contact controllers to produce planned behavior that explicitly leverages the inherent robustness of these controllers to fight unavoidable uncertainties in the models.