Abstract

Noisy observations coupled with nonlinear dynamics pose one of the biggestchallenges in robot motion planning. By decomposing the nonlinear dynamics intoa discrete set of local dynamics models, hybrid dynamics provide a natural wayto model nonlinear dynamics, especially in systems with sudden "jumps" in thedynamics, due to factors such as contacts. We propose a hierarchical POMDPplanner that develops locally optimal motion plans for hybrid dynamics models.The hierarchical planner first develops a high-level motion plan to sequencethe local dynamics models to be visited. The high-level plan is then convertedinto a detailed cost-optimized continuous state plan. This hierarchicalplanning approach results in a decomposition of the POMDP planning problem intosmaller sub-parts that can be solved with significantly lower computationalcosts. The ability to sequence the visitation of local dynamics models alsoprovides a powerful way to leverage the hybrid dynamics to reduce stateuncertainty. We evaluate the proposed planner for two navigation andlocalization tasks in simulated domains, as well as an assembly task with areal robotic manipulator.

 

Introduction (beta)

None

 

Conclusion (beta)

None