Recovery of Behaviors of Robots without Dynamics

For robots to function without humans on hand to fix them, they must be able to compensate for malfunctions that inhibit their ability to move all actuators.
The conventional choice would be to model the damaged robot, either by anticipating damage types and their effects on a known model, or by performing system identification on the damaged device. However, anticipating failure is fragile, and data-driven system ID of dynamics is expensive in time and robot wear-and-tear.
We instead suggest a method that leverages data from examples of successful motion to produce "behavior specifications" as constraints on outputs and states.
By collecting time-series data from (nearly arbitrary) observation functions on a working robot, we defined a behavior as a list of differential constraints.
The constraints, along with intrinsic constraints of physics, classify trajectories into those that yield a desired behavior, and those that do not.
For systems with many degrees-of-freedom, there are typically entire sub-manifolds of curves that meet the constraints.
Assuming that the damaged robot retains enough control freedom to re-enforce our constraints, we can design an input that restores a desired behavior.
We demonstrate our method on a crawler in simulation, and on a dynamic hexapod in hardware.