Detecting divergent robot behavior with multi-rigid body simulation and random sampling

The motion of robots and objects in our world is often highly dependent upon contact. When contact is expected but does not occur or when contact is not expected but does occur, robot behavior diverges from plan, often disastrously. We describe an approach that uses multi-rigid body simulation with contact toward predicting likely such behavioral divergences on real robots. Our approach, and others like it, could be applied to validation of robot behaviors, mechanism design, and even online planning. The standard approach to validating robot behavior is simulation followed by in situ testing. This approach does not inspire confidence, though, as simulations often fail to reflect real world behavior (we posit the cause below), and in situ testing is too tedious and slow to perform more than a few trials. This problem has instigated research into formal verification methods for robotics (e.g., [5], [7]), which appears promising; intense study is currently underway to scale these approaches to higher degree of freedom systems. We have been exploring an alternative path that is straightforward, easily implemented, and uses techniques already familiar to many roboticists (see [1], [4]) to bridge typical multi-body simulation tests and extensive testing on real robotic hardware. This approach uses multiple multi-body simulations with perturbations to initial conditions, modeling parameters, and sensory data toward capturing the likely span of situated robot behavior. Early experimental results indicate promise.