Search for: All records

The paper introduces DiSProD, an online planner developed forenvironments with probabilistic transitions in continuous state andaction spaces. DiSProD builds a symbolic graph that captures thedistribution of future trajectories, conditioned on a given policy,using independence assumptions and approximate propagation ofdistributions. The symbolic graph provides a differentiablerepresentation of the policy's value, enabling efficient gradient-basedoptimization for long-horizon search. The propagation of approximatedistributions can be seen as an aggregation of many trajectories, makingit well-suited for dealing with sparse rewards and stochasticenvironments. An extensive experimental evaluation compares DiSProD tostate-of-the-art planners in discrete-time planning and real-timecontrol of robotic systems. The proposed method improves over existingplanners in handling stochastic environments, sensitivity to searchdepth, sparsity of rewards, and large action spaces. Additionalreal-world experiments demonstrate that DiSProD can control groundvehicles and surface vessels to successfully navigate around obstacles. 

The paper provides a description of the ideas behind the DiSProD algorithm and system variants that participated and was the winner in the International Probabilistic Planning Competition, 2023.