Robust Control Chris Atkeson 5316 Combining modelbased control
Robust Control Chris Atkeson 5/3/16
Combining model-based control and predesigned policies: What do we do now? • We use a temporal hierarchy: • What to do next: behavior selection (seconds). The highest level involves selecting behaviors. Ignored but important. • Receding horizon control (seconds) Model-based trajectory generation and following. This level is very important for balance, but probably not for manipulation. • Actuator command generation (milliseconds) Model-based inverse dynamics. The importance of this level has been overstated in current work. • Robustifying policy (millisecond or faster) Joint-level PD control gains, for example. This importance of this level has been largely understated by academic robotics, and is crucial to good performance. Low-level fast policies are key contributors to robustness.
What’s next for robust control? • Much greater use of pre-designed policies. • Much greater emphasis on policy selection. • Much greater emphasis on policy learning. • Much greater emphasis on libraries of tuned and learned policies. • These libraries are shared across heterogeneous robots, so task vs. robot dependent learned information needs to be separated.
How do we design robust policies (and robust model-based control)? • We believe policy training approaches based on multiple models is a practical approach to policy training, and overall control system evaluation. • Current approaches to designing control policies for nonlinear systems are not robust. Linearization, for example, fails badly near singularities. Current optimal control approaches fail in this case, which is why our robots are all still walking with bent knees. • The multiple models need to cover the range of expected plant and task dynamics, and expected failure modes. • We believe the use of multiple-model simulation and during online optimal control is critical to reliability. • We believe massive amounts of redundant sensing are necessary for reliability. • We believe self monitoring outside the control loop is also necessary. We were one of the few DRC teams that actually took this seriously.
External Force Estimate 5
Finals Failed manipulation
Fall predictors (DRC Day 2 Drill Task) 7
- Slides: 7