An Efficient Environment Encoding for Trajectory Feasibility and Cost Predictions

Author: Zou, Sarah Jin

Year: 2022

Degree: Senior thesis (Major)

Advisor: Yue, Yisong

Committee Member: None, None

Option: Electrical Engineering; Information and Data Sciences

DOI: 10.7907/605s-5h91

Abstract

An efficient method to encoding a robot’s surrounding is important for the safety and scalability of path planning problems. I present a method for encoding occupied spaces in an environment via parametric geometry (e.g. circles). This encoding is utilized to train a cost and feasibility predicting neural network for path planning between two states. The circle encoding method uses either bounding circles or set cover to provide a lower dimension input of environments for neural network. The lower input dimension of the models allows for a smaller and faster model of similar performance. Feasibility and cost prediction neural networks enable lazy planning for trajectory calculation. While there are no safety-guarantees, the model I present can act as an heuristic to identify the best paths, amortizing a robot’s onboard computation. To train a cost and feasibility predicting model, I first generate multiple environments and calculate encoding circles for obstacle-occupied spaces. For each environment, I use planners to calculate multiple trajectories and store their feasibility and cost. I then train the feasibility and cost prediction models from this data . The feasibility model has 90% accuracy on train and test environments. The circle encoding method was compared with occupancy grid encoding method. When encoding methods were compared to predict occupied space, models with circle encoding generalized to unseen environments better than occupancy grid input models. The method of circle encoding I introduce can be utilized in other learning problems that use environment as input. The feasibility and cost predictions done in this work can be applied to bias tree growth or to other global planning methods.

Files