Counterfactual Multi-Agent Reinforcement Learning for Long- Horizon Medical Assistive Tasks with Dual-arm Robot.

Area of Science:

• Robotics
• Artificial Intelligence
• Healthcare Technology

Background:

• Dual-arm robots show promise for medical assistance but struggle with complex, sequential tasks.
• Multi-agent interactions and long-horizon planning pose significant challenges for current robot intelligence.

Purpose of the Study:

• To develop an advanced dual-arm robot intelligence for medical assistive tasks.
• To address challenges in multi-agent interactions and long-horizon actions using reinforcement learning.

Main Methods:

• Introduced Counterfactual Multi-Agent Demo Augmented Policy Gradient (COMA-DAPG), a novel multi-agent reinforcement learning approach.
• Integrated a counterfactual critic network and demonstration-augmented policy gradient (DAPG) with three reward functions.

Main Results:

• COMA-DAPG demonstrated superior performance compared to COMA and DAPG.
• Achieved over 25% improvement in average success rate across three long-horizon medical tasks.

Conclusions:

• COMA-DAPG effectively addresses credit assignment, gradient variance, and collision avoidance in dual-arm robotics.
• Enables precise and cooperative execution of complex medical tasks, boosting clinical reliability and efficiency.