Multi-Critic DDPG for Autonomous Indoor Drone Navigation

Abstract

This paper presents a novel Deep Deterministic Policy Gradient (DDPG)-based framework for autonomous in-door drone navigation in confined, unmapped environments. The proposed method employs a multi-layered DDPG control architecture coupled with a hybrid reward formulation that explicitly decomposes navigation into three complementary objectives: (i) obstacle avoidance via artificial repulsive potential fields, (ii) corridor centering through attraction-based alignment rewards, and (iii) adaptive pitch control to enable velocity modulation for efficient exploration. By jointly optimizing multiple control objectives, the drone achieves a higher degree of autonomy and robustness in complex indoor settings. We provide a detailed mathematical formulation of the control policy and reward structure, and validate the approach through extensive simulation experiments. The results demonstrate improved safety, stability, and exploration efficiency compared to baseline navigation strategies in cluttered indoor environments.

Ritabrata Chakraborty

ML Intern

Research Engineer specializing in robotics, computer vision, and autonomous systems. Currently developing automated data annotation solutions with foundation models for autonomous vehicles at Uber.