A Diffusion Model Augmented Indoor Robot Local Navigation Planner Based On Decision Transformer

Abstract

As the demand for indoor service robots grows, reliable local navigation techniques are crucial for their applications. Due to the speed limitations of interacting with the physical world, training robot navigation policies from pre-gathered of-fline datasets is becoming increasingly popular. However, training local navigation policies on offline datasets is not easy and faces two main challenges: (1) Myopia: Common assumptions based on Markov Decision Processes and limited local observations impose higher demands on navigation policies, making decision-making difficult for planners in complex scenarios. (2) Distribution Shift: Limited offline datasets only cover a portion of the state-action space, resulting in learned policies differing from behavioral policies and performing poorly in real world scenarios, while common non-generative data augmentation methods are not suitable for navigation scenarios. Therefore, we introduce a global information guided- diffusion model augmented local planner based on decision transformer. Our approach combines offline reinforcement learning and sequence modeling to train a local navigation algorithm, while a diffusion model guided by global information generates trajectories similar to the distribution of offline datasets but with different global paths, further improving navigation policies with synthetic data. Results in simulator demonstrate that this approach effectively improves the performance of various imitation learning algorithms on offline datasets and has good modeling capabilities for trajectories in multiple scenarios. Real-world tests also show that the algorithm can plan smooth, safe paths in complex scenarios, demonstrating excellent obstacle avoidance capabilities.