A Quantitative Comparison of Centralised and Distributed Reinforcement Learning-Based Control for Soft Robotic Arms

This paper proposes SoftGM, an octopus-inspired distributed control architecture for segmented soft robotic arms that learn to reach targets in contact-rich environments using online obstacle discovery without assuming full obstacle geometry at the start of each episode. SoftGM formulates each arm section as a cooperative agent and represents the arm-environment interaction as a graph. SoftGM uses a two-stage graph attention message passing scheme following a Centralised Training Decentralised Execution (CTDE) paradigm with a centralised critic and decentralised actor. We evaluate SoftGM in a Cosserat-rod simulator (PyElastica) across three tasks that increase the complexity of the environment: obstacle-free, structured obstacles, and a wall-with-hole scenario. Compared with six widely used MARL baselines (IDDPG, IPPO, ISAC, MADDPG, MAPPO, MASAC) under identical information content and training conditions, SoftGM matches strong CTDE methods in simpler settings and achieves the best performance in the wall-with-hole task. Robustness tests with observation noise, single-section actuation failure, and transient disturbances show that SoftGM maintains comparable performance under non-ideal simulated conditions while keeping control effort bounded, suggesting that selective contact-relevant information routing improves resilience in the tested settings.

Most reinforcement-learning (RL) controllers used in continuous control are architecturally centralized: observations are compressed into a single latent state from which both value estimates and actions are produced. Biological control systems are often organized differently. Insects, in particular, coordinate navigation, heading stabilization, memory, and context-dependent action selection through distributed circuits rather than a single monolithic controller. Motivated by this contrast, we study an RL policy architecture that decomposes control into interacting modules for sensory encoding, heading representation, sparse associative memory, recurrent command generation, and local motor control, with a learned arbitration mechanism that allocates motor authority across modules. The model is evaluated on a two-dimensional navigation task that require simultaneous food seeking, obstacle avoidance, and predator escape. In a six-seed predator-navigation experiment trained with Proximal Policy Optimization (PPO) for 75 updates, the modular policy achieves the strongest final mean performance among the tested controllers, with final episodic return $-2798.8\pm964.4$ versus $-3778.0\pm628.1$ for a centralized gated recurrent unit (GRU) and $-4727.5\pm772.5$ for a centralized multilayer perceptron (MLP). The modular policy also attains the lowest final value loss and stable PPO optimization statistics while driving module-assignment entropy to $0.0457\pm0.0244$, indicating highly selective control allocation. These results suggest that distributed control can serve as a useful inductive bias for RL problems involving dynamically competing behavioral objectives.

In animals such as elephants and octopuses, acquiring non-visual information about an object and physically engaging with it are inseparable processes mediated by rich, large-area interactions between compliant appendages and the environment. Soft robots provide a natural platform for translating this principle into engineered systems. Yet current robotic intelligence makes limited use of physical interaction, treating it primarily as a disturbance to be rejected or, at best, as a means of compensating for object misalignment. Here, we introduce a physical intelligence framework in which distributed compliant interactions jointly reveal task-relevant information and organize manipulation behavior. This results in an intrinsically partially observable problem: key task-relevant information is never measured directly, but must instead be inferred from the history of physical interactions. We propose a reinforcement-learning architecture that addresses this challenge by learning a memory-based control policy end-to-end. The key innovations making this possible are (i) a pretrained exploration policy that provides a reference for broad workspace exploration, (ii) joint optimization that integrates exploration and grasping objectives within a single recurrent policy, and (iii) a two-stage sim-to-real adaptation including observation mapping and policy fine-tuning. We demonstrate this principle through blind whole-arm grasping with a hybrid rigid-soft robotic arm that we equip with IMUs embedded directly within its compliant structure, providing its only source of proprioceptive sensing. The learned policy successfully identifies and grasps various objects by autonomously coordinating workspace exploration, object encounter and localization, inference of grasp-relevant properties, and stable whole-arm wrapping.