Sample-efficient and occlusion-robust reinforcement learning for robotic manipulation via multimodal fusion...

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Visual reinforcement learning (RL) advances bridge state-based and image-based training gaps.
• Robotic manipulation with occlusions remains a challenge for current visual RL methods.
• Tactile sensors offer solutions but are limited by cost and complexity.

Purpose of the Study:

• To develop a novel RL approach for robotic manipulation in occluded environments.
• To enhance sample efficiency and robustness without tactile feedback.
• To provide a cost-effective and scalable solution for real-world applications.

Main Methods:

• Introduced multimodal fusion dualization, optimizing actor and critic modules separately.
• Incorporated representation normalization techniques (LayerNorm, SimplexNorm) for stable training.
• Developed a visual RL approach without tactile sensors or prior knowledge.

Main Results:

• The proposed method effectively handles challenging robotic manipulation tasks with occlusions.
• Outperformed state-of-the-art visual RL and state-based RL in sample efficiency and task performance.
• Demonstrated robustness and scalability without tactile feedback or pre-trained representations.

Conclusions:

• The novel RL approach significantly improves performance in occluded robotic manipulation.
• Multimodal fusion dualization and representation normalization are key to success.
• The method presents a practical, cost-effective alternative to tactile sensing in robotics.