A neuromorphic control architecture for a biped robot

Michele Folgheraiter, A. Keldibek, Bauyrzhan Aubakir, Giuseppina Gini, Alessio Mauro Franchi, Matteo Bana

Research output: Contribution to journalArticle

Abstract

A neuromorphic control architecture is introduced to govern the motion of a lightweight humanoid robot. The reference trajectories necessary to perform stable gaits are generated by neural modules represented by Chaotic Recurrent Neural Networks CRNN organized in a hierarchical fashion. In the higher layer a body-coordination module generates the trajectories for the central parts of the robot body, in the middle layer the limb-coordination modules generate the Cartesian trajectories for the end effector of each limb, finally in the lower layer the limb modules control the position of the robot joints. Each neural module consists of a reservoir of N=200 leaky-integrator neurons randomly and sparsely connected with fixed synapses. The adaptation occurs in the synapses of readout units by online learning techniques like the delta rule and the Recursive Least Square algorithm RLS. It is demonstrated that the neural modules can learn and reproduce with enough accuracy the trajectories acquired from the simulation of a humanoid robot in V-REP software. With an optimal initialization of the reservoir connection matrix and by using a low computationally expensive learning algorithm such as the delta rule, Θ(N), the average of MSE over all lower limbs joints is in the order of 0.1. For the lower-limbs-coordination-module the MSE drops to 0.004 by using the more computational expensive RLS, Θ(N2). In case the neural module needs to learn how to adapt the trajectories according to a specific step length and frequency the MSE is 0.06. A comparison between different learning algorithms applied on the CRNN showed better performances by using RLS. This result is confirmed also by a direct comparison with a different neural architecture, the PCPG, however at the expense of a bigger computational complexity. A real test conducted on a small computational unit (Raspberry Pi2) demonstrated that the CRNN can be executed at a frequency of 142 Hz which suffices to feed a PID feedback control loop at the joint level.

Original languageEnglish
Article number103244
JournalRobotics and Autonomous Systems
Volume120
DOIs
Publication statusPublished - Oct 1 2019

Fingerprint

Biped Robot
Trajectories
Robots
Module
Trajectory
Learning algorithms
Humanoid Robot
Synapse
Recurrent neural networks
Three term control systems
Learning Algorithm
End effectors
Robot
Neurons
Feedback control
Connection Matrix
Computational complexity
Chaotic Neural Network
Unit
PID Control

Keywords

  • Biped robot
  • Chaotic Recurrent Neural Network
  • Humanoid robotics
  • Neurodynamics
  • Neuromorphic controller
  • Real time trajectory generation

ASJC Scopus subject areas

  • Control and Systems Engineering
  • Software
  • Mathematics(all)
  • Computer Science Applications

Cite this

A neuromorphic control architecture for a biped robot. / Folgheraiter, Michele; Keldibek, A.; Aubakir, Bauyrzhan; Gini, Giuseppina; Franchi, Alessio Mauro; Bana, Matteo.

In: Robotics and Autonomous Systems, Vol. 120, 103244, 01.10.2019.

Research output: Contribution to journalArticle

Folgheraiter, Michele ; Keldibek, A. ; Aubakir, Bauyrzhan ; Gini, Giuseppina ; Franchi, Alessio Mauro ; Bana, Matteo. / A neuromorphic control architecture for a biped robot. In: Robotics and Autonomous Systems. 2019 ; Vol. 120.
@article{fc33d0ee77cc49aab4792867c52af626,
title = "A neuromorphic control architecture for a biped robot",
abstract = "A neuromorphic control architecture is introduced to govern the motion of a lightweight humanoid robot. The reference trajectories necessary to perform stable gaits are generated by neural modules represented by Chaotic Recurrent Neural Networks CRNN organized in a hierarchical fashion. In the higher layer a body-coordination module generates the trajectories for the central parts of the robot body, in the middle layer the limb-coordination modules generate the Cartesian trajectories for the end effector of each limb, finally in the lower layer the limb modules control the position of the robot joints. Each neural module consists of a reservoir of N=200 leaky-integrator neurons randomly and sparsely connected with fixed synapses. The adaptation occurs in the synapses of readout units by online learning techniques like the delta rule and the Recursive Least Square algorithm RLS. It is demonstrated that the neural modules can learn and reproduce with enough accuracy the trajectories acquired from the simulation of a humanoid robot in V-REP software. With an optimal initialization of the reservoir connection matrix and by using a low computationally expensive learning algorithm such as the delta rule, Θ(N), the average of MSE over all lower limbs joints is in the order of 0.1. For the lower-limbs-coordination-module the MSE drops to 0.004 by using the more computational expensive RLS, Θ(N2). In case the neural module needs to learn how to adapt the trajectories according to a specific step length and frequency the MSE is 0.06. A comparison between different learning algorithms applied on the CRNN showed better performances by using RLS. This result is confirmed also by a direct comparison with a different neural architecture, the PCPG, however at the expense of a bigger computational complexity. A real test conducted on a small computational unit (Raspberry Pi2) demonstrated that the CRNN can be executed at a frequency of 142 Hz which suffices to feed a PID feedback control loop at the joint level.",
keywords = "Biped robot, Chaotic Recurrent Neural Network, Humanoid robotics, Neurodynamics, Neuromorphic controller, Real time trajectory generation",
author = "Michele Folgheraiter and A. Keldibek and Bauyrzhan Aubakir and Giuseppina Gini and Franchi, {Alessio Mauro} and Matteo Bana",
year = "2019",
month = "10",
day = "1",
doi = "10.1016/j.robot.2019.07.014",
language = "English",
volume = "120",
journal = "Robotics and Autonomous Systems",
issn = "0921-8890",
publisher = "Elsevier",

}

TY - JOUR

T1 - A neuromorphic control architecture for a biped robot

AU - Folgheraiter, Michele

AU - Keldibek, A.

AU - Aubakir, Bauyrzhan

AU - Gini, Giuseppina

AU - Franchi, Alessio Mauro

AU - Bana, Matteo

PY - 2019/10/1

Y1 - 2019/10/1

N2 - A neuromorphic control architecture is introduced to govern the motion of a lightweight humanoid robot. The reference trajectories necessary to perform stable gaits are generated by neural modules represented by Chaotic Recurrent Neural Networks CRNN organized in a hierarchical fashion. In the higher layer a body-coordination module generates the trajectories for the central parts of the robot body, in the middle layer the limb-coordination modules generate the Cartesian trajectories for the end effector of each limb, finally in the lower layer the limb modules control the position of the robot joints. Each neural module consists of a reservoir of N=200 leaky-integrator neurons randomly and sparsely connected with fixed synapses. The adaptation occurs in the synapses of readout units by online learning techniques like the delta rule and the Recursive Least Square algorithm RLS. It is demonstrated that the neural modules can learn and reproduce with enough accuracy the trajectories acquired from the simulation of a humanoid robot in V-REP software. With an optimal initialization of the reservoir connection matrix and by using a low computationally expensive learning algorithm such as the delta rule, Θ(N), the average of MSE over all lower limbs joints is in the order of 0.1. For the lower-limbs-coordination-module the MSE drops to 0.004 by using the more computational expensive RLS, Θ(N2). In case the neural module needs to learn how to adapt the trajectories according to a specific step length and frequency the MSE is 0.06. A comparison between different learning algorithms applied on the CRNN showed better performances by using RLS. This result is confirmed also by a direct comparison with a different neural architecture, the PCPG, however at the expense of a bigger computational complexity. A real test conducted on a small computational unit (Raspberry Pi2) demonstrated that the CRNN can be executed at a frequency of 142 Hz which suffices to feed a PID feedback control loop at the joint level.

AB - A neuromorphic control architecture is introduced to govern the motion of a lightweight humanoid robot. The reference trajectories necessary to perform stable gaits are generated by neural modules represented by Chaotic Recurrent Neural Networks CRNN organized in a hierarchical fashion. In the higher layer a body-coordination module generates the trajectories for the central parts of the robot body, in the middle layer the limb-coordination modules generate the Cartesian trajectories for the end effector of each limb, finally in the lower layer the limb modules control the position of the robot joints. Each neural module consists of a reservoir of N=200 leaky-integrator neurons randomly and sparsely connected with fixed synapses. The adaptation occurs in the synapses of readout units by online learning techniques like the delta rule and the Recursive Least Square algorithm RLS. It is demonstrated that the neural modules can learn and reproduce with enough accuracy the trajectories acquired from the simulation of a humanoid robot in V-REP software. With an optimal initialization of the reservoir connection matrix and by using a low computationally expensive learning algorithm such as the delta rule, Θ(N), the average of MSE over all lower limbs joints is in the order of 0.1. For the lower-limbs-coordination-module the MSE drops to 0.004 by using the more computational expensive RLS, Θ(N2). In case the neural module needs to learn how to adapt the trajectories according to a specific step length and frequency the MSE is 0.06. A comparison between different learning algorithms applied on the CRNN showed better performances by using RLS. This result is confirmed also by a direct comparison with a different neural architecture, the PCPG, however at the expense of a bigger computational complexity. A real test conducted on a small computational unit (Raspberry Pi2) demonstrated that the CRNN can be executed at a frequency of 142 Hz which suffices to feed a PID feedback control loop at the joint level.

KW - Biped robot

KW - Chaotic Recurrent Neural Network

KW - Humanoid robotics

KW - Neurodynamics

KW - Neuromorphic controller

KW - Real time trajectory generation

UR - http://www.scopus.com/inward/record.url?scp=85069895693&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85069895693&partnerID=8YFLogxK

U2 - 10.1016/j.robot.2019.07.014

DO - 10.1016/j.robot.2019.07.014

M3 - Article

AN - SCOPUS:85069895693

VL - 120

JO - Robotics and Autonomous Systems

JF - Robotics and Autonomous Systems

SN - 0921-8890

M1 - 103244

ER -