Towards Multi-Goal Navigation in the Real-World Using Reinforcement Learning

Abstract

Robots have proven to be highly effective in carrying out tasks requiring precise execution. However, when faced with real-world environments featuring sparse rewards and multiple goals, learning becomes a significant challenge, and Reinforcement Learning (RL) algorithms often struggle to acquire optimal policies. A common approach is to train in simulation environments and then fine-tune in the real world, although adapting to the real-world context remains challenging. This thesis introduces a novel method that addresses the challenge Sim2Real transfer of multi-goal navigation through reinforcement learning. The method utilizes object detectors as a pre-processing step to facilitate multi-goal navigation and transfer the learned knowledge to real-world scenarios. Empirical results demonstrate that the proposed method surpasses state-of-the-art baselines in terms of training time and overall performance, both in simulation environments and real-world settings. Although both methods achieve a perfect 100% success rate in simple environments, when it comes to single goal-based navigation, the proposed method demonstrates superior performance compared to the baseline in more challenging scenarios. Specifically, in complex environments and multi-goal scenarios within the AirSim environment, the proposed method surpasses the baseline by 18% and 5% respectively. As a proof of concept and to showcase its versatility, we also tested the approach in a low-fidelity environment, where it achieved similar levels of performance. To demonstrate the real-world implementation and validate the proposed method, we deployed it on a nano-drone named Crazyflie and a Jetbot with a front camera to conduct multi-goal navigation experiments. This practical deployment serves as a proof of concept, highlighting the effectiveness of the proposed method in real-world scenarios.