Abstract
The development of lightweight, reactive navigation systems for mobile edge robotics is a critical challenge in modern computer science. Traditional pathfinding relies on heavy, map-based computations which are often unsuited for hardware-constrained edge devices. This paper investigates the efficacy of deploying artificial neural networks to map raw sensor data directly to motor commands, bypassing traditional algorithmic pathfinding. We conduct a rigorous comparative study between two distinct machine learning paradigms: Behavioral Cloning (a supervised imitation learning approach) and Deep Q-Networks (a model-free reinforcement learning approach). Notably, to test the specific strengths of each paradigm, the supervised model navigated a continuous space with multiple randomized rectangular obstacles, while the reinforcement learning agent was tasked with finding a narrow pass-through corridor in a discrete-grid framework. Our findings indicate that while Behavioral Cloning is highly efficient at replicating hard-coded heuristic safety rules, the reinforcement learning agent—guided by potential-based reward shaping and DoubleDQN target updates—develops a superior capacity for complex spatial goal-seeking. Furthermore, this paper provides a deep mathematical analysis of observed training failures, specifically addressing the mechanics of reward plateaus and vanishing temporal-difference errors, and proposes concrete architectural improvements for future iterations.
Files
Software
- Language
- Python
- Development status
- Research prototype
References
- Thrun, S., Burgard, W., & Fox, D. (2005). Probabilistic Robotics. MIT Press.
- Pomerleau, D. A. (1989). ALVINN: An autonomous land vehicle in a neural network. Advances in Neural Information Processing Systems, 1.
- Mnih, V., Kavukcuoglu, K., Silver, D., et al. (2015). Human-level control through deep reinforcement learning. Nature, 518(7540), 529–533.
- Sutton, R. S., & Barto, A. G. (2018). Reinforcement Learning: An Introduction. MIT Press.
- Ng, A. Y., Harada, D., & Russell, S. (1999). Policy invariance under reward transformations: Theory and application to reward shaping. ICML, 99, 278–287.
- Van Hasselt, H., Guez, A., & Silver, D. (2016). Deep reinforcement learning with double Q-learning. Proceedings of the AAAI Conference on Artificial Intelligence, 30(1).