[Paper Review] Scalability in Neural Control of Musculoskeletal Robots

Christoph Richter, Sören Jentzsch+7 more|arXiv (Cornell University)|Jan 19, 2016

Advanced Memory and Neural Computing30 references18 citations

TL;DR

This paper presents a scalable neural control framework for musculoskeletal robots by integrating the Myorobotics hardware platform with the SpiNNaker neuromorphic computing system. It demonstrates real-time, low-power control of up to 12 joints using a closed-loop cerebellar model, enabling extensible, brain-inspired control for human-scale anthropomimetic robots.

ABSTRACT

Anthropomimetic robots are robots that sense, behave, interact and feel like humans. By this definition, anthropomimetic robots require human-like physical hardware and actuation, but also brain-like control and sensing. The most self-evident realization to meet those requirements would be a human-like musculoskeletal robot with a brain-like neural controller. While both musculoskeletal robotic hardware and neural control software have existed for decades, a scalable approach that could be used to build and control an anthropomimetic human-scale robot has not been demonstrated yet. Combining Myorobotics, a framework for musculoskeletal robot development, with SpiNNaker, a neuromorphic computing platform, we present the proof-of-principle of a system that can scale to dozens of neurally-controlled, physically compliant joints. At its core, it implements a closed-loop cerebellar model which provides real-time low-level neural control at minimal power consumption and maximal extensibility: higher-order (e.g., cortical) neural networks and neuromorphic sensors like silicon-retinae or -cochleae can naturally be incorporated.

Motivation & Objective

• To address the lack of scalable, real-time, and energy-efficient neural control for human-scale anthropomimetic musculoskeletal robots.
• To overcome limitations in existing robotic control platforms, such as inflexible neuromorphic systems and non-scalable neural simulators.
• To enable integration of higher-order neural networks and neuromorphic sensors (e.g., silicon retina, cochlea) into a unified, extensible control architecture.
• To demonstrate a proof-of-concept system that supports real-time, closed-loop neural control with minimal power consumption and high extensibility.

Proposed method

• Integration of the Myorobotics musculoskeletal hardware platform with the SpiNNaker neuromorphic computing platform for real-time neural simulation.
• Implementation of a closed-loop cerebellar model on SpiNNaker to provide low-level, real-time neural control of compliant joints.
• Use of event-based sensory interfaces (e.g., AER for vision and audition) that natively map to SpiNNaker’s spike-based packet communication.
• Leveraging open-source neural simulation frameworks like PyNN and Nengo to port and run complex neural network models on SpiNNaker with minimal modification.
• Employing a distributed FlexRay-based communication architecture to scale joint control beyond the native 12-joint limit via multiple SpiNN-IO boards.
• Utilizing hall-effect encoders to sense tendon displacement and calculate muscle force from known spring constants and routing geometry.

Experimental results

Research questions

• RQ1 Can a neuromorphic platform like SpiNNaker provide real-time, low-power, and scalable neural control for a large number of compliant musculoskeletal joints?
• RQ2 How can brain-inspired neural controllers be effectively integrated with physical musculoskeletal robotic hardware to enable natural, human-like behavior?
• RQ3 To what extent can existing neuromorphic sensors (e.g., silicon retina, cochlea) be natively interfaced with a neuromorphic control system for real-time sensory feedback?
• RQ4 Can higher-level neural networks (e.g., cortical models) be seamlessly incorporated into a scalable neural control stack built on SpiNNaker and Myorobotics?
• RQ5 What are the practical limits of scalability in neural control for anthropomimetic robots, and how can they be overcome using distributed neuromorphic architectures?

Key findings

• The system successfully achieved real-time control of a musculoskeletal joint at a 500 Hz update rate using a simulated cerebellar model on SpiNNaker.
• Up to 12 joints (24 actuators) could be controlled simultaneously using a single SpiNN-IO board with a dedicated FlexRay controller, with scalability extended via multiple boards.
• Event-based sensory systems such as silicon retinae and cochleae were natively compatible with the SpiNNaker platform through address-event representation (AER) mapping.
• Neural network models specified in PyNN or Nengo could be ported to SpiNNaker with minimal modifications, enabling rapid integration of complex brain-inspired controllers.
• The framework supports extensible integration of higher-order neural networks and learning rules, enabling future development of full brain-like control systems.
• The combination of Myorobotics and SpiNNaker enables a scalable, low-power, and extensible platform suitable for building human-scale anthropomimetic robots with brain-like control.