International Conference on Electronics Technology (ICET)
Compute-Constrained Variational Quantum Reinforcement Learning for Energy-Aware Autonomous Uav Navigation in Dynamic Environments
Muhammad Rashid Majeed, Peiyun Zhang, P.G. Balis, Mst Jannatul Kobra, Owahedur Rahman, Zamadder Md Iqbal Hossain
Abstract
Applying quantum-improved reinforcement learning to autonomous UAV systems is inherently faced with the challenge of limited energy and computing resources onboard, especially in dynamic and wind-turbulent scenarios. To overcome this problem, we present a compute-adaptive quantum-classical control approach that executes quantum policies within tight runtime budgets while maintaining navigation performance and safety. The approach presents a two-policy paradigm that combines a high-capacity variational quantum policy with a lightweight classical proxy policy, both managed by a learned budget controller that modulates quantum policy usage in real time via a skip probability. The approach is trained using the PPO algorithm on a stochastic 3D UAV navigation task and shows that the proposed scheduler effectively tracks a target quantum budget, increasing the training-time skip probability from about 0.35 to close to 1.0 while maintaining compute-adaptive fluctuations in quantum usage between .0.1 and 1.0, depending on the episode complexity. Notably, learning is stable even under high environmental stochasticity, with episodic rewards ranging from about -700 to -100, and the moving average of the last 50 episodes converges and remains within -300 to -340 over more than 6,000 training episodes, demonstrating stable convergence with compute-aware switching. In addition, the trajectory rollouts validate the effectiveness of obstacle-aware navigation towards the goal in 3D space, with smooth energy consumption curves that gradually decrease from normalised energy 1.0 to around 0.73 before termination, without mid-trajectory energy spikes. Failure mode analysis aggregated across months also shows that the most common failure cause is boundary violation, at around 77%, followed by energy depletion at 18%, and collisions at 5%, which are realistic safety and energy trade-offs rather than instability. Finally, experiments in generalisation across months show strong peak robustness, with best-case success rates at around 0.95, 0.97, and 0.90 under different wind regimes, although final policy performance shows sensitivity. In conclusion, the above experiments demonstrate that quantum reinforcement learning can be stabilised and deployed within computational constraints, with clear scope for improving boundary violations and generalisation under extreme environmental uncertainty.