Dikole, Realeboga GBaloyi, Andrew ASekopa, Teboho L2026-03-132026-03-132025-12979-8-3315-5876-5DOI: 10.1109/ICCR67607.2025.11372069http://hdl.handle.net/10204/14763This research explores a vision-based Deep Deterministic Policy Gradient (DDPG) algorithm for robotic manipulations within the ABB IRB120 robot arm scenario. The robot is trained to perform reach tasks based on RGB image observations within a simulated OpenAI Gym environment. Performance is determined by comparing the ABB robot’s performance against that of the Franka Emika Panda robot using the same training regime. The study also contrasts image-based observations with traditional sensor data, with different learning efficiency and convergence. The experiment results show that the vision-based DDPG algorithm has a 70% success rate, confirming it can learn control policies directly from visual observations. In addition, the ABB robot exhibits more stable learning behaviour than the Panda robot. These findings support the use of vision-driven reinforcement learning for precise and dynamic robotic control.AbstractenDeep reinforcement LearningVision-based learningRoboticsConference PresentationN/A