Voronoi Tessellation for Efficient Sampling in Gaussian Process-Based Robotic Motion Planning

On-line motion planning in dynamically changing environments poses a significant challenge in the design of autonomous robotic system. Conventional methods often require intricate design choices, while modern deep reinforcement learning (DRL) approaches demand vast amounts of robot motion data. Gaussian process (GP) regression-based imitation learning approaches address such issues by harnessing the GP's data-efficient learning capabilities to infer generalized policies from a limited number of demonstrations, which can intuitively be generated by human operators. GP-based methods, however, are limited in data scalability as computation becomes cubically expensive as the amount of learned data increases. This issue is addressed by proposing Voronoi tessellation sampling, a novel data sampling strategy for learning GP-based robotic motion planning, where spatial correlation between input features and the output of the trajectory prediction model is exploited to select the data to be learned that are informative yet learnable by the model. Where the baseline is set by an imitation learning framework that uses GP regression to infer trajectories that learns policies optimized via a stochastic, reward-based optimization algorithm, experimental results demonstrate that the proposed method can learn optimal policies spanning over all of feature space using fewer data compared to the baseline method.