Self-learning of Robot Arm using a Modified Soft-Actor-Critic (SAC) Approach: Complex 6D Pose Brick Grasping Test

Previous studies have not addressed robot arm grasping for objects with complex 6D poses using reinforcement learning (RL) when the pose is known beforehand. This is crucial for automation in architecture and related fields. To address this gap, we aimed to achieve precise grasping of objects with complex poses. We employed the Soft Actor-Critic (SAC) baseline and its two variants (using median and mean Q values) to demonstrate the limitations of standard RL algorithms and the effectiveness of modified algorithms for this complex task. The trained models were evaluated using statistical metrics, including average episodic rewards, steps, and success rates. Our results indicate that both SAC variants effectively trained the agent to achieve the target 6D pose, with the mean variant performing slightly better than the median variant. Test rollouts demonstrate that the mean variant of the SAC algorithm exhibits superior performance in attaining positions and poses that are more proximate to the target object compared to the median variant.