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

TL;DR

This work tackles agile manipulation for robotic arms operating in dynamic, unknown environments. It introduces CBMC-V3, a CNS-inspired framework built on fully SNNs with five modules, three hierarchical control levels, and two information pathways to orchestrate high-speed, robust control. The cerebellum learns gravity patterns, the thalamus dynamically weights these patterns by load, the brainstem tunes PD gains online, and the spinal cord provides fast high-frequency feedback, all integrated with cortex-based trajectory storage. Experimental validation in simulation and on a real 7-DOF arm demonstrates superior agility and adaptation over industrial controllers, especially under varying payloads and higher-speed motions.

Abstract

As robotic arm applications extend beyond industrial settings into service-oriented sectors such as catering, household and retail, existing control algorithms struggle to achieve the agile manipulation required for complex environments with dynamic trajectories, unpredictable interactions, and diverse objects. This paper presents a biomimetic control framework based on Spiking Neural Networks (SNNs), inspired by the human Central Nervous System (CNS), to achieve agile control in such environments. The proposed framework features five control modules (cerebral cortex, cerebellum, thalamus, brainstem, and spinal cord), three hierarchical control levels (first-order, second-order, and third-order), and two information pathways (ascending and descending). Each module is fully implemented using SNN. The spinal cord module uses spike encoding and Leaky Integrate-and-Fire (LIF) neurons for feedback control. The brainstem module employs a network of LIF and non-spiking LIF neurons to dynamically adjust spinal cord parameters via reinforcement learning. The thalamus module similarly employs a network of LIF and non-spiking LIF neurons to adjust the cerebellum's torque outputs via reinforcement learning. The cerebellum module, which provides feedfoward gravity compensation torques, uses a recurrent SNN to learn the robotic arm's dynamics through regression. The framework is validated both in simulation and on real-world robotic arm platform under various loads and trajectories. Results demonstrate that our method outperforms the industrial-grade position control in manipulation agility.

Figures (8)

• Figure 1: Overview of the human CNS and the proposed framework structure. (a) Human CNS relating to motion control. Sensory information and motor commands are transmitted through three orders of neurons along ascending and descending pathways. Sensory signals travel from peripheral receptors to the spinal cord, brainstem, and ultimately the cerebellum and cerebral cortex, while motor commands descend from the cerebral cortex and cerebellum through the thalamus and spinal cord to the muscles. The spinal cord can independently generate reflexive responses via reflex arcs, enabling rapid reactions to stimuli LLOBERA2023190. The brainstem contributes to movement termination, reorientation, and coordination with the spinal cord merel2019hierarchicalannurev:/content/journals/10.1146/annurev-neuro-082321-025137. The thalamus functions as a hub for motor coordination, multimodal perception, and sensorimotor integration LLOBERA2023190Halassa_2022. The cerebellum regulates body movements, balance, and posture by integrating sensory inputs and refining motor commands LLOBERA2023190. The cerebral cortex generates decisions and motion trajectories. (b) Schematic diagram of our proposed framework, featuring five modules, three control levels and two pathways as mentioned in (a). (c) Neural network design of each module, utilizing three types of spiking neurons: input fiber (grey), LIF neuron (yellow) and non-spiking LIF neuron (green).
• Figure 2: (a) Structure of one network unit in the cerebellum module during training. (b) Structure of cerebellum network during prediction.
• Figure 3: Structure of thalamus module.
• Figure 4: Average feedback control coefficients of each joint across 3 trajectory tracking tasks.
• Figure 5: (a) Average tracking error using single mode and two modes across three trajectories. (b) Average weight generated by the thalamus module across three trajectories.