Enhancing stochastic mobility prediction models for robust planetary navigation on unstructured terrain

Abstract

Motion planning for planetary rovers must consider control uncertainty in order to maintain the safety of the platform during navigation. Modelling such control uncertainty is challenging on unstructured terrain, and especially on deformable terrain, due to the complex interaction between the platform and its environment. In this thesis, we propose to enhance stochastic transition models for planning, where the outcomes of the control actions are learnt from experience and represented statistically using probability density functions. These transition models that capture control uncertainty are known as Stochastic Mobility Prediction Models (SMPM). Rovers may traverse a mixture of rigid and deformable terrain. However, current SMPMs are only capable of estimating one dimension of control uncertainty in rigid terrain. We propose to enhance the SMPM by Learning from Exteroception, a training method that relies on sample executions of motion primitives on representative terrain and the corresponding platform configurations collected along the executed path. This method enables the estimation of the outcome of future control actions on deformable terrain. The SMPM is further enhanced by using multi-output Gaussian process regression by simultaneously considering the correlation between multiple dimensions of uncertainty. The enhanced SMPM is integrated into a Markov decision process framework and dynamic programming is used to construct a control policy for navigation to a goal region in a terrain map. We consider both rigid and deformable terrain, consisting of uneven ground, small rocks, and non-traversable rocks. Over 300 experimental trials are reported using a planetary rover platform in a Mars-analogue environment. Our results demonstrate increased path safety and reliability by the improved traversal cost and actions executed; due to the SMPM improvement in predicting control action outcomes.