Loading...
Guidance and Control of Multi-Agent Free Flying Robots Using Deep Reinforcement Learning
Delavar, Ali | 2025
120
Viewed
- Type of Document: M.Sc. Thesis
- Language: Farsi
- Document No: 58663 (45)
- University: Sharif University of Technology
- Department: Management and Economics
- Advisor(s): Kiani, Maryam
- Abstract:
- Free-flying robots are critical for supporting human efforts during space missions. Recent advancements in multi-agent robotics, enabled by Multi-Agent Deep Reinforce- ment Learning (MADRL), have further enhanced the efficiency and autonomy of these systems. Multi-agent actor-critic methods with Centralized Training and Decentralized Execution (CTDE) enables robots for decentralized coordination in a continuous action space by addressing the problems of non-stationary and credit assignment in multi- agent environments. Notably, Multi-Agent Proximal Policy Optimization (MAPPO) has demonstrated superior performance compared to other methods in these environ- ments. Integrating the attention layer inside the actor-critic’s network architecture mitigates the partial observability issue in multi-agent settings and enhances the colli- sion avoidance with other agents and obstacles. These innovations are transforming how autonomous multi-agents perform tasks, enabling high-reliability operations inside and outside spacecraft. This Thesis employs MAPPO integrated with the attention mech- anism for motion planning and collision avoidance of homogeneous free-flying robots in 3d space
- Keywords:
- Deep Reinforcement Learning ; Attention Mechanism ; Multi-Agent Reinforcement Learning ; Proximal Policy Optimization Algorithm ; Multi-Agent Robotics ; Free-Flying Robots ; Centralized Training and Decentralized Execution (CTDE)
-
محتواي کتاب
- view
