Dynamic Obstacle Avoidance in Autonomous Robot Navigation Using Deep Reinforcement Learning
| dc.contributor.author | Parmash A. | |
| dc.date.accessioned | 2026-09-09T05:53:51Z | |
| dc.date.available | 2026-09-09T05:53:51Z | |
| dc.date.issued | 2026 | |
| dc.description.abstract | Existing indoor navigation systems for autonomous robots have proven their success in geometric mapping and avoidance of obstacles in the local environment, but they lack semantic reasoning capabilities; thus, they cannot comprehend instructions issued in natural language. In contrast, utilizing state-of-the-art vision and language foundation models in the robot control loop would create too much computational delay, making the implementation impractical for the purpose of reactive execution of tasks. This thesis attempts to fill this gap through a two-stage hierarchy. During the initial phase, a mobile robot performs an exploration task within the unknown environment by employing an SLAM pipeline based on LiDAR, while the robot’s on-board VLM performs object detection to build a structured, small global 3D semantic database. During the second phase, the robot takes up any natural-language instructions (for example, “go to the bookshelf”), extracts the coordinates of the target from its memory, and performs autonomous navigation. Here, the navigation process is entirely mapless and does not involve any VLM inference cost overhead. Control at the lower level is done by the robot’s DRL-based policy trained using TD3 algorithms. Assessed under ROS-integrated Gazebo simulation, the proposed architecture attains an accuracy rate of 85% in navigation with a lightweight Moondream VLM, substantially superior to cloud-based GPT-4o mini versions and classical approaches. The trained policy for depth proves to have excellent generalization performance with respect to unknown objects, dense obstacles, and even multiple rooms, as well as a success rate of 80% in evading moving obstacles without the need for further training. Achieving a processing rate of 10 Hz on board, the framework manages to bypass foundation model delay constraints. | |
| dc.identifier.citation | Parmash A / Dynamic Obstacle Avoidance in Autonomous Robot Navigation Using Deep Reinforcement Learning / SDU University / Department of Computer Science | |
| dc.identifier.uri | https://repository.sdu.edu.kz/handle/123456789/2657 | |
| dc.language.iso | en | |
| dc.publisher | SDU University | |
| dc.subject | Deep Reinforcement Learning | |
| dc.subject | TD3 | |
| dc.subject | Vision-Language Models (VLMs) | |
| dc.subject | Mapless Semantic Navigation | |
| dc.title | Dynamic Obstacle Avoidance in Autonomous Robot Navigation Using Deep Reinforcement Learning | |
| dc.type | Thesis |