pith. sign in

arxiv: 2606.11034 · v1 · pith:IXI4LAEQnew · submitted 2026-06-09 · 💻 cs.RO · cs.NE

A Spiking Neural Architecture for Coordinating Arm and Locomotor Control

Lea Steffen , Kathryn Simone , Graeme Damberger , Travis DeWolf , Hudson Ly , Chris Eliasmith This is my paper

Pith reviewed 2026-06-27 13:10 UTC · model grok-4.3

classification 💻 cs.RO cs.NE
keywords spiking neural networkshumanoid robot controlbipedal locomotionarm controlbasal ganglianeural engineering frameworksemantic pointer architectureneuromorphic hardware
0
0 comments X

The pith

A spiking neural architecture coordinates force-based arm control and bipedal locomotion on a simulated humanoid robot using a basal ganglia model for task switching.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents an integrated spiking neural controller that handles both arm movements and walking in one system on a full humanoid. It uses biologically inspired spiking models to implement the controls, with a spiking basal ganglia model selecting between actions. The system is tested in simulation for reaching targets, drawing with digits, following paths while walking, and switching tasks. This matters because spiking networks can run efficiently on low-power neuromorphic chips, potentially allowing robots to perform multiple motor tasks without separate controllers. The work shows successful integration where previous systems treated locomotion and arm control separately.

Core claim

The paper demonstrates a spiking neural architecture that integrates force-based arm control and bipedal locomotion on a simulated full-scale humanoid platform. High-level action selection is handled by a biologically grounded spiking basal ganglia model. Validation through co-simulation demonstrates successful target reaching, continuous digit drawing, path-following locomotion, and switching between walking and arm control via basal ganglia disinhibition.

What carries the argument

A spiking basal ganglia model that mediates switching between locomotor and arm control tasks through disinhibition, integrated with spiking implementations of the motor controllers.

Load-bearing premise

The co-simulation between the neural simulator and the robot physics engine accurately represents real-world dynamics, sensor noise, and the basal ganglia's role in task switching.

What would settle it

Failure of the system to maintain stable walking or accurate reaching when the controller is transferred from simulation to a physical humanoid robot would indicate the claim does not hold.

Figures

Figures reproduced from arXiv: 2606.11034 by Chris Eliasmith, Graeme Damberger, Hudson Ly, Kathryn Simone, Lea Steffen, Travis DeWolf.

Figure 1
Figure 1. Figure 1: The simulation snapshots show walking (green) and arm control for drawing the digit ‘8’ (blue). Below are shown the spike trains of the basal ganglia (strD1, strD2, STN and GPi), which are used to coordinate these actions. Dashed lines link the robot states to their corresponding neural activity. Abstract Spiking Neural Networks (SNNs) coupled with neuromorphic hard￾ware offer energy-efficient solutions fo… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the integrated spiking motor control system. The basal ganglia (brown) performs action selection, gating either the arm controller (blue) or locomotion controller (green). Dashed lines indicate proprioceptive feedback; solid lines indicate control signals. based framework is used to control the right arm of the Unitree H1, which has 4 DOF. It includes three biologically grounded modules: the so… view at source ↗
Figure 4
Figure 4. Figure 4: Drawing the digit ‘5’. (a) Resulting 2D end-effector trajectory (red) accurately tracing the reference template for the digit "5" (dashed). (b) Isaac Sim visualization of the Unitree H1 actively performing the drawing motion. digit (dashed). In Figure 4b, the Unitree H1 robot performs the com￾manded drawing motion within the Isaac Sim physics environment. For all arm experiments (see [PITH_FULL_IMAGE:figu… view at source ↗
Figure 3
Figure 3. Figure 3: Reach-to-target task, if end-effector surpasses 4 cm threshold a new target is generated. (a) Continuous tracking error over time. The turquoise solid line represents the Euclidean distance between the robotic end-effector and the randomized spatial targets. (b) Neural spike raster plots and corresponding decoded variables for each opera￾tional axis (𝑋, 𝑌, 𝑍). The physical end-effector trajectory (green), … view at source ↗
Figure 6
Figure 6. Figure 6: Spiking continuous action selection in the basal ganglia. (Top) Input command values overlaid with the spikes of striatal D1 (strD1) neurons, representing the excitatory drive for the ‘WALK’ and ‘DRAW’ actions over time. (Bottom) Action selection outputs overlaid on the globus pallidus internus (GPi) spiking raster. Actions are selected through disinhibition, so high output activity indicates the action is… view at source ↗
read the original abstract

Spiking Neural Networks (SNNs) coupled with neuromorphic hardware offer energy-efficient solutions for humanoid robot control. However, existing SNN-based motor control systems address bipedal locomotion and arm control in isolation, leaving integrated control of both unaddressed. We present a spiking architecture that coordinates force-based arm control and bipedal locomotion in a simulated humanoid, using the Neural Engineering Framework (NEF) and Semantic Pointer Architecture (SPA). High-level action selection between locomotor and arm control is mediated by a biologically grounded spiking basal ganglia model. We validate the system through co-simulation of Nengo, for the neural control, and Isaac Sim, demonstrating successful target reaching, continuous digit drawing, path-following locomotion, and finally, switching between walking and arm control via basal ganglia disinhibition. To our knowledge, this is the first integrated spiking controller to combine bipedal locomotion and arm control on a full-scale humanoid platform. The full spike-based implementation enables future deployment on low-power neuromorphic hardware.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper presents a spiking neural architecture using the Neural Engineering Framework (NEF), Semantic Pointer Architecture (SPA), and a biologically grounded spiking basal ganglia model to coordinate force-based arm control and bipedal locomotion on a simulated full-scale humanoid robot. It demonstrates target reaching, digit drawing, path-following locomotion, and switching between tasks via basal ganglia disinhibition through co-simulation with Nengo and Isaac Sim, claiming to be the first such integrated spiking controller.

Significance. If the integration and switching results hold under more rigorous testing, this would represent a meaningful engineering contribution as the first integrated spiking controller combining bipedal locomotion and arm control on a full-scale humanoid platform. The work builds explicitly on established prior frameworks (NEF, SPA, basal ganglia model) without introducing new fitted parameters, providing a concrete path toward neuromorphic hardware deployment.

major comments (2)
  1. [Abstract and validation description] Abstract and validation description: the central claim of successful coordination and clean switching (mediated by the basal ganglia model) rests on qualitative statements of 'successful' demonstrations in co-simulation; no quantitative metrics such as tracking error, success rates, interference measures between primitives, or baseline comparisons are reported, which is load-bearing for asserting that the architecture achieves integrated control without cross-talk.
  2. [Validation experiments] Validation experiments: the co-simulation is presented as sufficient validation, but no analysis is provided of robustness to unmodeled effects (contact forces, actuator delays, sensor noise) on the disinhibition signals or task interference; this directly affects the claim that the basal ganglia model ensures clean mediation between locomotion and arm control.
minor comments (1)
  1. [Abstract] The novelty claim ('to our knowledge, this is the first...') in the abstract would be strengthened by an explicit comparison to prior isolated locomotion or arm SNN controllers in a dedicated related-work section.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which highlight important aspects of validation for our integrated spiking controller. We address each major comment below and indicate where revisions will be made to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract and validation description] Abstract and validation description: the central claim of successful coordination and clean switching (mediated by the basal ganglia model) rests on qualitative statements of 'successful' demonstrations in co-simulation; no quantitative metrics such as tracking error, success rates, interference measures between primitives, or baseline comparisons are reported, which is load-bearing for asserting that the architecture achieves integrated control without cross-talk.

    Authors: We agree that the current manuscript relies on qualitative descriptions of task success and switching. To address this, the revised version will include quantitative metrics computed from the existing co-simulations, such as end-effector position tracking error during reaching and digit drawing, path deviation error during locomotion, task completion success rates, and a measure of cross-talk (e.g., unintended movement in the non-selected subsystem during switching). Where data permits, we will also add simple baseline comparisons to non-integrated or non-spiking variants. revision: yes

  2. Referee: [Validation experiments] Validation experiments: the co-simulation is presented as sufficient validation, but no analysis is provided of robustness to unmodeled effects (contact forces, actuator delays, sensor noise) on the disinhibition signals or task interference; this directly affects the claim that the basal ganglia model ensures clean mediation between locomotion and arm control.

    Authors: This is a valid point regarding the strength of the claims about clean mediation. In revision we will add a dedicated discussion subsection on robustness, grounded in the biological properties of the basal ganglia model (e.g., its known tolerance to noise in disinhibition). We will also report results from additional targeted simulations that introduce controlled levels of contact force variation, actuator delay, and sensor noise to assess effects on switching reliability and task interference. A full parametric robustness study across all possible unmodeled effects is beyond the scope of this proof-of-concept paper, but the added analysis will directly support the mediation claim. revision: partial

Circularity Check

0 steps flagged

No circularity: engineering demonstration using established frameworks

full rationale

The paper describes an implementation that combines NEF, SPA, and a prior basal ganglia model for task switching, then shows results in Nengo-Isaac Sim co-simulation. No equations, fitted parameters, or predictions are presented that reduce to the inputs by construction. Novelty is asserted as the first integrated spiking controller on a full-scale humanoid, but this is an empirical claim about the system, not a derivation. Self-citations to the underlying frameworks are standard and not load-bearing for any mathematical result. The work is self-contained as an engineering demonstration.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Based solely on abstract; no explicit free parameters or invented entities described. Relies on established frameworks assumed to transfer to this integrated use case.

axioms (2)
  • domain assumption The Neural Engineering Framework and Semantic Pointer Architecture are suitable for force-based motor control and high-level action representation.
    Invoked as the foundation for the spiking architecture.
  • domain assumption A biologically grounded spiking basal ganglia model can mediate high-level action selection between locomotion and arm control via disinhibition.
    Used for switching behavior without further justification in abstract.

pith-pipeline@v0.9.1-grok · 5717 in / 1398 out tokens · 48702 ms · 2026-06-27T13:10:35.135703+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

23 extracted references · 16 canonical work pages

  1. [1]

    Stewart, Daniel Rasmussen, Xuan Choo, Aaron Voelker, and Chris Eliasmith

    Trevor Bekolay, James Bergstra, Eric Hunsberger, Travis DeWolf, Terrence C. Stewart, Daniel Rasmussen, Xuan Choo, Aaron Voelker, and Chris Eliasmith. 2014. Nengo: a Python tool for building large-scale functional brain models.Frontiers in Neuroinfor- matics7 (Jan. 2014). doi:10.3389/fninf.2013.00048

    work page doi:10.3389/fninf.2013.00048 2014

  2. [2]

    Djork-Arné Clevert, Thomas Unterthiner, and Sepp Hochreiter. 2015. Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs). arXiv: Learning(Nov. 2015). https://www.semanticscholar.org/paper/Fast- and-Accurate-Deep-Network-Learning-by-Linear-Clevert-Unterthiner/ f63e917638553414526a0cc8550de4ad2d83fe7a

    2015

  3. [3]

    de Azambuja, A

    R. de Azambuja, A. Cangelosi, and S. V. Adams. 2016. Diverse, noisy and parallel: a New Spiking Neural Network approach for humanoid robot control. In2016 Interna- tional Joint Conference on Neural Networks (IJCNN). 1134–1142. doi:10.1109/IJCNN. 2016.7727325 ISSN: 2161-4407

    work page doi:10.1109/ijcnn 2016

  4. [4]

    Travis DeWolf, Kinjal Patel, Pawel Jaworski, Roxana Leontie, Joe Hays, and Chris Eliasmith. 2023. Neuromorphic control of a simulated 7-DOF arm using Loihi.Neu- romorphic Computing and Engineering3, 1 (Feb. 2023), 014007. doi:10.1088/2634- 4386/acb286

    work page doi:10.1088/2634- 2023

  5. [5]

    Stewart, Jean-Jacques Slotine, and Chris Eliasmith

    Travis DeWolf, Terrence C. Stewart, Jean-Jacques Slotine, and Chris Eliasmith. 2016. A spiking neural model of adaptive arm control.Proceedings of the Royal Society B: Biological Sciences283, 1843 (Nov. 2016), 20162134. doi:10.1098/rspb.2016.2134

    work page doi:10.1098/rspb.2016.2134 2016

  6. [6]

    and Chetwynd, Amanda G

    Chris Eliasmith. 2013.How to Build a Brain: A Neural Architecture for Biological Cognition. Oxford University Press. doi:10.1093/acprof:oso/9780199794546.001.0001

    work page doi:10.1093/acprof:oso/9780199794546.001.0001 2013

  7. [7]

    Anderson

    Chris Eliasmith and Charles H. Anderson. 2003.Neural Engineering: Computation, Representation, and Dynamics in Neurobiological Systems. MIT Press. Google-Books- ID: J6jz9s4kbfIC

    2003

  8. [8]

    and Choo, Xuan and Bekolay, Trevor and DeWolf, Travis and Tang, Yichuan and Rasmussen, Daniel , year = 2012, month = nov, journal =

    Chris Eliasmith, Terrence C. Stewart, Xuan Choo, Trevor Bekolay, Travis DeWolf, Yichuan Tang, and Daniel Rasmussen. 2012. A Large-Scale Model of the Functioning Brain.Science338, 6111 (Nov. 2012), 1202–1205. doi:10.1126/science.1225266

    work page doi:10.1126/science.1225266 2012

  9. [9]

    Eric Hunsberger and Chris Eliasmith. 2016. Training Spiking Deep Networks for Neuromorphic Hardware. (2016). doi:10.13140/RG.2.2.10967.06566 arXiv:1611.05141 [cs]

    work page doi:10.13140/rg.2.2.10967.06566 2016

  10. [10]

    isaac-sim. [n. d.]. H1FlatTerrainPolicy. https://github.com/isaac- sim/IsaacSim/blob/develop/source/extensions/isaacsim.robot.policy.examples/ isaacsim/robot/policy/examples/robots/h1.py

  11. [11]

    Xiaoyang Jiang, Qiang Zhang, Jingkai Sun, Jiahang Cao, Jingtong Ma, and Renjing Xu. 2025. Fully Spiking Neural Network for Legged Robots. InICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). 1–5. doi:10.1109/ICASSP49660.2025.10890793

    work page doi:10.1109/icassp49660.2025.10890793 2025

  12. [12]

    Ashwin Sanjay Lele, Yan Fang, Justin Ting, and Arijit Raychowdhury. 2020. Learning to Walk: Bio-Mimetic Hexapod Locomotion via Reinforcement-Based Spiking Central Pattern Generation.IEEE Journal on Emerging and Selected Topics in Circuits and Systems10, 4 (Dec. 2020), 536–545. doi:10.1109/JETCAS.2020.3033135

    work page doi:10.1109/jetcas.2020.3033135 2020

  13. [13]

    2022.Human Standing Dynamic Balancing and Elementary Biolog- ical Signal Processing

    Bálint Petró. 2022.Human Standing Dynamic Balancing and Elementary Biolog- ical Signal Processing. Ph.D. Budapest University of Technology and Economics (Hungary), Hungary. https://www.proquest.com/docview/3064677953/abstract/ FF9885C8D9C24159PQ/1

    arXiv 2022

  14. [14]

    Michmizos

    Ioannis Polykretis, Guangzhi Tang, Praveenram Balachandar, and Konstantinos P. Michmizos. 2022. A Spiking Neural Network Mimics the Oculomotor System to Control a Biomimetic Robotic Head Without Learning on a Neuromorphic Hardware. IEEE Transactions on Medical Robotics and Bionics4, 2 (May 2022), 520–529. doi:10. 1109/TMRB.2022.3155278

    arXiv 2022

  15. [15]

    Rostro-Gonzalez, P

    H. Rostro-Gonzalez, P. A. Cerna-Garcia, G. Trejo-Caballero, C. H. Garcia-Capulin, M. A. Ibarra-Manzano, J. G. Avina-Cervantes, and C. Torres-Huitzil. 2015. A CPG system based on spiking neurons for hexapod robot locomotion.Neurocomputing 170 (Dec. 2015), 47–54. doi:10.1016/j.neucom.2015.03.090

    work page doi:10.1016/j.neucom.2015.03.090 2015

  16. [16]

    Terrence C Stewart, Xuan Choo, and Chris Eliasmith. [n. d.]. Dynamic Behaviour of a Spiking Model of Action Selection in the Basal Ganglia. ([n. d.])

  17. [17]

    Tokey Tahmid, Mark Gates, Piotr Luszczek, Catherine Schuman, Catherine Schuman, and Catherine Schuman. 2026. SpikeRL: A Scalable and Energy-efficient Framework for Deep Spiking Reinforcement Learning. doi:10.1109/ICONS69015.2025.00033

    work page doi:10.1109/icons69015.2025.00033 2026

  18. [18]

    Michmizos

    Guangzhi Tang, Neelesh Kumar, Raymond Yoo, and Konstantinos P. Michmizos. 2020. Deep Reinforcement Learning with Population-Coded Spiking Neural Network for Continuous Control. doi:10.48550/arXiv.2010.09635 arXiv:2010.09635 [cs]

    work page doi:10.48550/arxiv.2010.09635 2020

  19. [19]

    Camilo Vasquez Tieck, Lea Steffen, Jacques Kaiser, Daniel Reichard, Arne Roennau, and Ruediger Dillmann

    J. Camilo Vasquez Tieck, Lea Steffen, Jacques Kaiser, Daniel Reichard, Arne Roennau, and Ruediger Dillmann. 2019. Combining Motor Primitives for Perception Driven Target Reaching With Spiking Neurons:.International Journal of Cognitive Informatics and Natural Intelligence13, 1 (Jan. 2019), 1–12. doi:10.4018/IJCINI.2019010101

    work page doi:10.4018/ijcini.2019010101 2019

  20. [20]

    Camilo Vasquez Tieck, Lea Steffen, Jacques Kaiser, Arne Roennau, and Rudiger Dillmann

    J. Camilo Vasquez Tieck, Lea Steffen, Jacques Kaiser, Arne Roennau, and Rudiger Dillmann. 2018. Controlling a Robot Arm for Target Reaching without Planning Using Spiking Neurons. In2018 IEEE 17th International Conference on Cognitive Informatics & Cognitive Computing (ICCI*CC). IEEE, Berkeley, CA, 111–116. doi:10.1109/ICCI- CC.2018.8482049

    work page doi:10.1109/icci- 2018

  21. [21]

    Camilo Vasquez Tieck, Lea Steffen, Jacques Kaiser, Arne Roennau, and Rudiger Dill- mann

    J. Camilo Vasquez Tieck, Lea Steffen, Jacques Kaiser, Arne Roennau, and Rudiger Dill- mann. 2018. Multi-Modal Motion Activation for Robot Control Using Spiking Neurons. In2018 7th IEEE International Conference on Biomedical Robotics and Biomechatronics (Biorob). IEEE, Enschede, 291–298. doi:10.1109/BIOROB.2018.8487786

    work page doi:10.1109/biorob.2018.8487786 2018

  22. [22]

    George-Iulian Uleru, Mircea Hulea, and Vasile-Ion Manta. 2022. Using Hebbian Learning for Training Spiking Neural Networks to Control Fingers of Robotic Hands. International Journal of Humanoid Robotics19, 06 (Dec. 2022), 2250024. doi:10.1142/ S0219843622500244

    2022

  23. [23]

    2025.Proxy Target: Bridging the Gap Between Discrete Spiking Neural Networks and Continuous Control

    Zijie Xu, Tong Bu, Zecheng Hao, Jianhao Ding, and Zhaofei yu. 2025.Proxy Target: Bridging the Gap Between Discrete Spiking Neural Networks and Continuous Control. doi:10.48550/arXiv.2505.24161

    work page doi:10.48550/arxiv.2505.24161 2025