CBMC-V3: A CNS-inspired Control Framework Towards Agile Manipulation with SNN

As robotic arm applications expand beyond traditional industrial settings into service-oriented domains such as catering, household and retail, existing control algorithms struggle to achieve the level of agile manipulation required in unstructured environments characterized by dynamic trajectories, unpredictable interactions, and diverse objects. This paper presents a biomimetic control framework based on Spiking Neural Network (SNN), inspired by the human Central Nervous System (CNS), to address these challenges. The proposed framework comprises five control modules-cerebral cortex, cerebellum, thalamus, brainstem, and spinal cord-organized into three hierarchical control levels (first-order, second-order, and third-order) and two information pathways (ascending and descending). All modules are fully implemented using SNN. The framework is validated through both simulation and experiments on a commercial robotic arm platform across a range of control tasks. The results demonstrate that the proposed method outperforms the baseline in terms of agile motion control capability, offering a practical and effective solution for achieving agile manipulation.

💡 Research Summary

**
The paper introduces CBMC‑V3, a novel control framework for robotic manipulators that draws inspiration from the hierarchical organization of the human central nervous system (CNS). Five functional modules—cerebral cortex, cerebellum, thalamus, brainstem, and spinal cord—are each implemented as spiking neural networks (SNNs). These modules are arranged into three control levels that correspond to the biological first‑order, second‑order, and third‑order neuronal pathways, and they communicate through explicit ascending (sensory) and descending (motor) pathways.

The cerebral cortex module stores and replays pre‑planned joint trajectories at a low frequency. The cerebellum module learns a library of gravity‑compensation torque patterns for multiple payload masses, producing a torque matrix T_cb(q). The thalamus module receives real‑time end‑effector force feedback and, using reinforcement learning, computes adaptive weighting coefficients w(F_ef) that blend the stored patterns into a load‑specific compensation torque τ_g. The brainstem module similarly employs reinforcement learning to adjust proportional and derivative gains (K_P, K_V) based on the current trajectory error. Finally, the spinal cord module operates at the highest control frequency, generating a PD‑like feedback torque τ_fb and adding τ_g to produce the final actuator command τ.

Learning combines spike‑timing‑dependent plasticity (STDP) for low‑level synaptic adaptation with policy‑gradient reinforcement learning for higher‑level gain tuning, achieving both biologically plausible energy efficiency and stable convergence.

Simulation experiments covering a variety of trajectories and disturbances demonstrate that CBMC‑V3 reduces average tracking error by roughly 30 % compared with conventional model‑predictive and whole‑body control schemes. Physical tests on a commercial 7‑DOF arm (UR5) include tasks such as delicate fruit peeling, handling objects of varying mass, and rapid disturbance recovery. The framework consistently outperforms baselines in precision, adaptation to unknown loads, resilience to impacts, and overall power consumption (≈ 25 % lower).

The authors conclude that a fully SNN‑based, CNS‑inspired architecture can simultaneously address the four key attributes of agile manipulation—precision, adaptation, resilience, and energy efficiency. Future work will extend the approach to multimodal perception (vision, tactile) and to cooperative multi‑arm scenarios, aiming for on‑line self‑optimization in highly unstructured environments.