I have a working grid-based environment search-and-rescue drone path planner that combines A* with a reinforcement-learning layer. It already performs reasonably well, but I want it to outperform current state-of-the-art solutions in three concrete areas: increased efficiency (shorter mission time), better obstacle avoidance, and superior route optimisation. On top of that, the reinforcement-learning module must introduce a truly novel power-saving approach so that flight endurance is extended without compromising performance. The existing code is clean and fully documented; you will receive the repo plus a small set of benchmark scenarios. Your mandate is to redesign the RL reward structure, refine or replace the current A* heuristics, and integrate any other innovative techniques you deem fit—multi-agent coordination, dynamic re-planning, hybrid meta-heuristics, etc.—as long as the final system: • Clears every benchmark faster than the baseline while consuming less simulated energy • Shows measurable gains in obstacle avoidance (fewer collisions or near misses) • Produces routes with lower total cost than the baseline and at least one leading academic reference implementation Provide the updated source, a short technical report with comparative plots, and a reproducible test script so I can verify the results locally. The project is self-contained; no field deployment is required at this stage, only simulation. If you have experience with RL libraries (e.g., TensorFlow, PyTorch), heuristic search tuning, or drone power modelling, I would love to see how you can push this algorithm beyond current limits.