CMAX++ : Leveraging Experience in Planning and Execution using Inaccurate Models

Given access to accurate dynamical models, modern planning approaches are effective in computing feasible and optimal plans for repetitive robotic tasks. However, it is difficult model the true dynamics of real world before execution, especially tasks requiring interactions with objects whose parameters unknown. A recent approach, CMAX, tackles this problem by adapting planner online during execution bias resulting away from inaccurately modeled regions. while being provably guaranteed reach goal, requires strong assumptions on accuracy used fails improve quality solution over repetitions same task. In paper we propose CMAX++, an approach that leverages real-world experience successive a CMAX++ achieves integrating model-free learning using acquired model-based potentially inaccurate model. We provide provable guarantees completeness asymptotic convergence path cost as number increases. also shown outperform baselines simulated including 3D mobile robot navigation where track friction incorrectly modeled, 7D pick-and-place task mass object unknown leading discrepancy between dynamics.