Deep RL based Obstacle Avoidance for UAVs with Time Varying sensor Bias

Aditya Kurande¹    Bhaskar Joshi²    Harikumar Kandath²   

¹ BITS Pilani K K Birla Goa Campus, Goa    ² Robotics Research Center, IIIT Hyderabad, India   

Advancements in Deep Reinforcement Learning (DRL) have shown significant promise for the development of self-directed Unmanned Aerial Vehicles (UAVs). Our research focuses on examining the impact of time-varying noise such as that present in Inertial Measurement Unit (IMU) sensors on the effi-cacy of DRL-based waypoint navigation and obstacle avoidance in UAVs. Sensor noise affects localization and obstacle detection, and is assumed to follow a Gaussian probability distribution with an unknown non-zero time varying mean and variance. In this study, we consider environment with static obstacles where noise exhibits a time-varying bias, a characteristic commonly associated with IMU sensors. We evaluate the effectiveness of a DRL agent, trained using the Proximal Policy Optimization (PPO) technique, in the presence of such noise.