Octopus-inspired Distributed Control for Soft Robotic Arms: A Graph Neural Network-Based Attention Policy with Environmental Interaction
A new graph neural network architecture enables soft robotic arms to navigate complex obstacles without a global map.
A research team including Linxin Hou, Qirui Wu, and Cecilia Laschi has published a paper on arXiv introducing SoftGM, a novel control architecture for soft robotic arms. Inspired by the decentralized, resilient coordination of an octopus's tentacles, SoftGM treats each segment of a soft arm as an independent agent. These agents cooperate using a two-stage graph attention neural network, which models the arm's interaction with its environment as a graph. This allows the system to learn effective reaching and manipulation policies through a Centralised Training, Decentralised Execution (CTDE) paradigm, where a central critic guides training but each segment acts independently during execution.
The system was rigorously evaluated in a physics simulator (PyElastica) across three increasingly complex tasks: obstacle-free reaching, navigating structured obstacles, and a challenging wall-with-hole scenario. SoftGM was benchmarked against six established multi-agent reinforcement learning (MARL) algorithms, including MADDPG and MAPPO. While performance was comparable in simpler settings, SoftGM achieved the best success rate in the most complex wall task. Crucially, the architecture demonstrated significant robustness, maintaining performance despite added observation noise, simulated actuator failures in single sections, and transient disturbances. This resilience stems from its graph-based attention mechanism, which dynamically routes only contact-relevant information between segments, keeping control effort bounded and enabling adaptive behavior in unpredictable environments.
- Uses a Graph Neural Network (GNN) with attention for decentralized control of soft arm segments as cooperative agents.
- Outperformed 6 major MARL baselines (including MADDPG, MAPPO) in complex 'wall-with-hole' navigation tasks in simulation.
- Showed robust performance against observation noise, single-section failures, and disturbances due to selective information routing.
Why It Matters
Enables more adaptive and resilient soft robots for real-world applications like search-and-rescue, surgery, and handling delicate objects in cluttered spaces.