Masterstudent i Reinforcement Learning hos Toyota Material Handling i Mjölby

Develop and evaluate a reinforcement learning approach for generating test scenarios for autonomous forklift software.

Work with an existing simulation and testing environment and explore how reinforcement learning can interact with it.

Define relevant scenario parameters and investigate how the learning process can be guided toward challenging or interesting operating conditions.

Compare reinforcement learning with simpler scenario-generation approaches, such as random testing and other search or optimization methods.

Analyze discovered failure scenarios and evaluate test coverage, failure discovery rate, and testing efficiency.

Investigate whether failure scenarios discovered in simulation can be reproduced and validated on a physical autonomous truck.

The thesis will be carried out with support from Toyota Material Handling, including guidance on the existing testing environment, available data, and autopilot system. As this is an exploratory research project, well-founded partial results are considered valuable outcomes, and the goal is not to deliver a production-ready solution.

Validation of autonomous forklift software requires testing across a large number of operating conditions. Warehouse layouts, obstacles, pallet locations, vehicle configurations, traffic, and other environmental factors create a scenario space that is difficult to cover through manually designed test cases alone.

The purpose of this thesis is to investigate whether reinforcement learning can be used to intelligently generate test scenarios. Rather than controlling the forklift, the RL agent selects the conditions under which the autonomous forklift is tested. Based on the outcome of previous tests, the agent learns which areas of the scenario space are more likely to expose failures, unexpected behavior, or challenging edge cases.

The approach will be evaluated against simpler testing strategies to determine whether reinforcement learning can improve failure discovery, test coverage, or testing efficiency. An important outcome will be an objective assessment of when RL provides value and when simpler approaches may be sufficient.