The Optimized SAC Algorithm Enables Efficient Motion Control of the Tensegrity Spherical Robot

Spherical tensegrity robots exhibit superior environmental adaptability and application potential by integrating the advantages of tensegrity structures. This paper focuses on motion control of a six-bar tensegrity spherical robot driven by propellers, addressing challenges in dynamic modeling and limitations of traditional control methods. A novel SAC reinforcement learning algorithm is proposed, incorporating two key improvements: a structure-based double-layer attention mechanism and an improved prioritized experience replay (PER). The attention mechanism decomposes the state space according to the robot's geometric structure, enabling focused intra-group rod coordination and efficient inter-group coupling modeling. The improved PER integrates SimHashbased similarity penalties and adaptive priority truncation to enhance sample diversity and training stability. Simulation experiments in MuJoCo demonstrate significant advantages in reward acquisition, network loss optimization, and training efficiency. Comparative studies against DDPG, A2C, and PPO show superior performance, while ablation experiments validate the contribution of each component. The robot successfully navigates both flat and uneven terrains, demonstrating excellent motion performance and environmental adaptability with positioning accuracy.