A Feedback-Driven Learning Framework for Adaptive Neural Motion Planner

This paper presents a feedback-driven learning framework for adaptive neural motion planners to address the challenges of motion planning in high-dimensional and complex environments. Recently, learning-based motion planning frameworks have shown promise for providing efficient solutions; however, these methods face significant generalization challenges when encountering unseen tasks, resulting in a substantial decline in planning success rates. To tackle this, our framework integrates continual learning techniques with feedback loops, enabling the planner to autonomously adapt to new tasks while mitigating catastrophic forgetting. We propose a novel sample selection strategy that combines the GEM gradient projection method with an adaptive greedy algorithm. This algorithm, guided by beta distribution sampling, dynamically adjusts episodic memory updates based on task success rates, ensuring that the samples remain diverse and representative. Extensive experiments in complex environments validate the superior adaptability, robustness, and efficiency of the framework compared to existing methods.