Learning based modular framework for environment-adaptive planning in exploration

Abstract

Autonomous planning has spawned a number of different solutions in robotics research, using different paradigms and strategies, both generalized and specific to certain problems, representations, and environments.

Thus, we pose the hypothesis that the hyperparameters for best performance are dependent on the type of environment determined by its “clutteredness”. Usually, these hyperparameters are manually tuned by trial and error for a general estimate of the type of environments likely to be encountered by the agent, not accounting for different environment types within the same map. While the planner may operate within optimal bounds, the absolute best performance of the planner on every map in the dataset is unguaranteed.

This problem becomes evident and significant in exploration tasks, where the agent could encounter different environments during the same task. For exploration tasks, especially search-and-rescue, efficient navigation is critical, which is to have the highest or best performance in every portion of the map explored while looking for objects of interest. To address this issue this work presents a modular framework to utilize the experience of the underlying structure and the corresponding planning strategies performances of prior environments to predict the unknown map and adapt the planning strategy for maintaining high performance throughout the current exploration task.

In this thesis, we evaluate this framework using A* with the inflation factor hyperparameter being adapted to the environment in the task to see improvement over static hyperparameters. This framework can be extended to other heuristic-based planners and even sets of planners where the adaptation is in which the planner is being used in the environment rather than the hyperparameters of the planner model.