Decoding Neural Population Dynamics through Robotic Analog

Introduction

• Rotational neural dynamics appear during movement, but their contribution to generating movement remained unclear.Source: Introduction · p. 2
• Motor-system redundancy and monitoring only a small fraction of neurons hinder linking neural dynamics to motor output.Source: Introduction · p. 2
• Most ANN studies reconstructed simplified behavioral or neuromuscular signals, often overlooking the gap between execution and motion outcomes.Source: Introduction · p. 2
• Biomechanical simulations addressed this gap, but the absence of real-world implementation left motor output disconnected from realistic sensory information.Source: Introduction · p. 3
• The proposed robotic analog combines distributed artificial muscles, multimodal sensors, and reinforcement-learning-trained ANN control to study neural dynamics in physical movement.Source: Introduction · p. 3
• The authors use this task-driven embodied model to examine rotational dynamics, sensory effects on neural energy, and motor-learning “Eureka” moments.Source: Introduction · p. 3

Robotic analog of the biological motor system

• The arm uses self-sensing artificial-muscle lengths, joint-angle IMUs, and vision-based end-effector position as sensory feedback.Source: Robotic analog of the biological motor… · p. 3
• Controllers were trained in simulation using reinforcement learning, with training randomization introduced to support transfer to the physical arm.Source: Robotic analog of the biological motor… · pp. 3–4
• An end-to-end ANN actor receives sensory feedback and directly outputs stimulating currents to individual artificial muscles.Source: Robotic analog of the biological motor… · pp. 3–4
• Without physical-robot fine-tuning, controllers achieved mean reaching error below 5 mm across conditions and generalized to arbitrary trajectories.Source: Robotic analog of the biological motor… · p. 4
• The controller traced complex trajectories, including writing the university emblem, while maintaining autonomous, adaptive movement.Source: Robotic analog of the biological motor… · pp. 4, 6

Similarities between robotic and biological motor control

• After unexpected muscle damage, the robot maintained reaching accuracy below 5 mm without retraining or adjustment, compensating with other functional muscles.The tested damage scenarios included failures of one muscle, two muscles in different layers, or three muscles in the same layer.Source: Similarities between robotic and… · p. 6
• Population activity of 512 RNN neurons exhibited two distinct phases, preparatory and execution; their first PCs explained 20.39+3.10% of variance and were nearly orthogonal, with angle close to 90°.The separation extends across multiple dimensions rather than being confined to a single axis.Source: Similarities between robotic and… · pp. 6–7
• Execution and preparatory activity were largely segregated but not strictly orthogonal: each placed more than 0.5 of its variance in the other subspace’s null space, while overlap remained below 0.5.The pattern held across different feedback combinations and formed distinct but partially overlapping neural manifolds.Source: Similarities between robotic and… · p. 7
• Neural population variance dropped sharply at the state transition, while preparatory activity remained generally output-null and specified upcoming movement without affecting ongoing muscle activity.This organization parallels the preparatory role described for motor-cortical dynamics.Source: Similarities between robotic and… · p. 7
• A local linear approximation captured neural-state evolution (R2: 0.9–1.0), with dynamics shifting from output-null directions during preparation toward output-potent directions during execution.The shift reflects reorientation of neural dynamics rather than a global increase in gain.Source: Similarities between robotic and… · p. 7
• A small number of dominant dynamical modes supported performance, and those modes were not strongly aligned with output-potent directions, indicating computations within internal dynamics before output transformation.Removing the top-ranked modes substantially degraded performance.Source: Similarities between robotic and… · p. 7
• Neural–kinematic alignment was low-dimensional, concentrated in a few leading modes, and epoch-specific; tensor maximum-entropy surrogate data did not reproduce this organization.Neural population activity also shifted globally between preparatory and execution states in shared state space.Source: Similarities between robotic and… · p. 8
• During execution, the controller exhibited spiral-like rotational neural activity across single-point reaching and cycling tasks and across feedback combinations.The rotational dynamics occurred in the 512-neuron population and were consistent across the tested controller variations.Source: Similarities between robotic and… · p. 8

• RNN activity aligned near 90° with its temporal evolution, and its rotational-model fit exceeded TME-based surrogates preserving data covariance; rotations were confined to RNN layers, not input or muscle outputs.Source: Similarities between robotic and… · p. 10
• Across identical task conditions and random initializations, RNNs showed stronger, more coherent rotations and better behavioral performance than MLPs, ruling out seed-specific effects.Source: Similarities between robotic and… · pp. 10–11
• Increasing control channels from 3 to 6 to 12 progressively improved rotational consistency and movement accuracy, with 12 channels producing the most concentrated rotations around 90° and most effective control.Source: Similarities between robotic and… · p. 11
• Rotational dynamics remained consistent in simulations and real-world experiments, indicating that sensory feedback delays and noise did not influence the rotations.Source: Similarities between robotic and… · p. 11

Decoding rotational neural dynamics

• Rotational dynamics generate movement near 90° to the reaching direction, supporting orthogonal maneuvers for trajectory adjustment.Source: Decoding rotational neural dynamics · pp. 11–12, Figure 3 · p. 12
• As jPC1-reconstructed movement approached orthogonality, controllers showed higher corresponding movement velocity.Source: Decoding rotational neural dynamics · p. 13
• With three control channels, reconstructed movement deviated from 90°; six or twelve channels moved the angle toward 90° while improving velocity and accuracy.Source: Decoding rotational neural dynamics · p. 13
• CCA projections of primate neural activity reconstructed trajectories aligned with actual reaches, while jPC1 projections showed pronounced orthogonal oscillations.Source: Decoding rotational neural dynamics · p. 13
• The robotic analog showed the same jPC1-associated pattern across sensory-feedback combinations and control strategies.Source: Decoding rotational neural dynamics · pp. 13–14
• The output-potent subspace captured most target-related signal (R2: 0.8–0.9), while the jPC subspace explained little variance (R2 < 0.15) and was geometrically distinct.Source: Decoding rotational neural dynamics · p. 14
• Pruning output connections of rotation-dominant units increased reaching error more than intact or randomly pruned networks.Source: Decoding rotational neural dynamics · p. 14

Neural energy under different feedback and actuation combinations

• Neural energy was time-averaged squared neural activity; added sensory input increased input-layer energy but reduced RNN-layer energy.The layer-specific pattern contrasts with the expectation that processing more sensory inputs would raise overall neural energy.Source: Neural energy under different feedback… · p. 16
• As sensory channels increased, RNN activity converged faster to a stable low-energy state, reducing overall neural energy across simulated and physical arms.Input-layer energy increased as RNN energy decreased.Source: Neural energy under different feedback… · p. 16
• Across 15 seeds, vision-only models favored sparse muscle use, while Len+Ang+Vis produced more distributed, smoother activation; distal muscles activated earlier and more strongly.Structural trends persisted despite variation in individual muscle contributions.Source: Neural energy under different feedback… · p. 16
• Increasing control channels lowered neural energy while maintaining comparable muscle energy and enabling more frequent, variable actuation adjustments.The pattern was observed in simulation and physical experiments.Source: Neural energy under different feedback… · p. 17
• Both sensory feedback and motor redundancy reduced neural energy without increasing actuation-energy demand.Increasing control channels also produced more frequent, variable actuation adjustments.Source: Neural energy under different feedback… · pp. 16–17

Eureka moments during the learning process

• The robotic analog enabled controller learning to be unrolled, addressing the difficulty of tracking full neural development in biological systems.The paper presents this as a way to examine how neural networks develop motor-control capabilities.Source: Eureka moments during the learning… · p. 17
• Near episode 2,000, accurate reaching emerged alongside stable rotational dynamics and a sharp decline in network sparsity.PCA-explained variance and rotation radius also jumped and then plateaued.Source: Eureka moments during the learning… · p. 19
• At the same training point, preparatory and execution subspaces became nearly orthogonal, and neural and muscle actuation energies shifted.The angular separation moved toward 90°.Source: Eureka moments during the learning… · p. 19
• Coordinated changes in sparsity, population dynamics, and behavior recurred across tested architectures, feedback combinations, and random initializations, with slight variation in transition timing.Both RNN and MLP architectures exhibited transitions.Source: Eureka moments during the learning… · p. 19
• RNN layers developed stronger rotational dynamics and achieved higher performance than MLP architectures.Both architectures nevertheless exhibited learning transitions.Source: Eureka moments during the learning… · p. 19
• The detailed neural mechanisms underlying the observed reorganization remain to be investigated.The proposed interaction between neural dynamics and body mechanics is described as a possible mechanism.Source: Eureka moments during the learning… · pp. 19–20

Discussion

• The robotic analog combines artificial muscles, multimodal sensors, and an ANN controller trained by reinforcement learning, achieving accurate, damage-robust movement and motor-cortex-like rotational dynamics.The authors describe it as a paradigm beyond in vivo and in silico approaches.Source: Discussion · p. 20
• The platform offers physical embodiment and computational transparency for hypothesis-driven experiments and fine-grained analyses that would be challenging or infeasible in vivo.The discussion presents it as a generalizable paradigm for investigating principles shaping motor control.Source: Discussion · pp. 20–22
• Rotational dynamics drive orthogonal maneuvers for online direction correction, helping compensate for errors while maintaining stability.The discussion describes these as effective trajectory adjustments.Source: Discussion · p. 20
• Preparatory and execution subspaces emerge without pre-designed structure and persist across real and simulated settings despite feedback delays.The authors argue this persistence supports functional roles for the dynamics rather than sensory delays driving them.Source: Discussion · p. 20
• Behavioral performance is strongly coupled to neural rotations across input-output configurations and learning, with causal perturbation analyses further supporting their functional relevance.The discussion also notes correspondence with cerebellar findings on orthogonal Purkinje-cell vectors.Source: Discussion · pp. 20–21
• Richer feedback and more control channels can reduce neural-controller computational burden while maintaining motor performance, as they disambiguate sensorimotor mappings.The authors connect redundant state and action choices with more efficient solutions for state estimation and motor planning.Source: Discussion · p. 21
• At the 2000th training episode, rotations, preparatory states, and accurate movements emerged alongside an abrupt sparsity drop; the transition recurred across seeds, architectures, and feedback combinations.The discussion interprets this robust transition as an insight-like reorganization of internal representations.Source: Discussion · p. 21
• The authors suggest that bio-inspired orthogonal dynamics and control redundancy may inform more adaptive, stable, and energy-efficient robots, while the current framework focuses on online closed-loop control.Adding memory consolidation and replay is described as a future direction.Source: Discussion · p. 22

Disclosure of AI-Assisted technologies

• Reinforcement learning was used as a core research method, while no large language models or other AI-assisted tools were used to write or present the manuscript.Source: Disclosure of AI-Assisted technologies · p. 22

Fabrication of the physical model

• The physical model used square-knot LCE muscles with conductive steel-fiber coatings, quaternion-orientation IMUs, and triangular rigid segments with latex membranes for elastic restoring force.Source: Fabrication of the physical model · pp. 23–24

Assembly of the simulated model

• The simulated model combined an exponential muscle resistance–length relationship, an Ogden hyperelastic model extended with temperature and rate terms, and a latex membrane represented by three linear springs.Source: Assembly of the simulated model · pp. 24–25

Reinforcement learning model training

• Randomly sampling new targets after successful reaches and rewarding the number reached per episode encouraged movement strategies that generalized beyond specific trajectories and initial conditions.The unified task varied initial states and target positions.Source: Reinforcement learning model training · p. 27
• Randomizing artificial-muscle parameters across episodes and adding Gaussian noise to exteroceptive feedback supported policy transfer across physical and sensor variations without real-world fine-tuning or retraining.Muscle parameters were sampled from uniform distributions centered on experimentally measured values.Source: Reinforcement learning model training · p. 27
• Policies used combinations of visual end-effector position, resistance-based muscle-length feedback, and IMU joint orientation; proprioception-only controllers used a predictor to infer missing visual feedback.The tested combinations included vision-only, muscle-length-only, muscle length plus joint angle, muscle length plus vision, joint angle plus vision, and all three modalities.Source: Reinforcement learning model training · pp. 27–28
• Predictors learned end-effector position from rollouts of a pretrained policy, were fine-tuned on a small set of real-robot motion data, and used an LSTM architecture trained to minimize mean squared error.Stochastic policy actions during data collection were used to improve predictor robustness to realistic control variation.Source: Reinforcement learning model training · p. 29

Data acquisition

• Camera calibration verification showed that a single calibration remained accurate for several months.Source: Data acquisition · p. 30
• Robot feedback included motion-capture end-effector tracking at 50 Hz, resistance-based muscle-length estimates, and IMU quaternion measurements converted to joint angles.Resistance was measured from circuit voltage and controller output current; the passage does not specify a sampling rate for the IMU data.Source: Data acquisition · p. 30

Data analysis

• For jPCA, neural activity across time and conditions was reduced to the top 30 principal components, differentiated numerically, and used to fit a linear dynamical model.Source: Data analysis · pp. 30–31
• Tensor maximum entropy surrogates preserved covariances across neurons, conditions, and time while disrupting interactions, and were analyzed with jPCA over 100 iterations.The surrogate analysis tested whether rotational dynamics could arise from neuronal tuning and smoothness alone.Source: Data analysis · p. 31
• jPCA reduced neural activity with PCA, fitted a skew-symmetric model to identify rotational planes, and assessed fit by the variance in temporal derivatives explained.The model predicts the neural-state derivative from the skew-symmetric matrix and neural state.Source: Data analysis · pp. 31–32
• The leading jPC1 projection defined a Rotation state that was replicated across 60 time steps as an RNN signal, whose action-driven robot movement was termed jPC1-reconstructed.The synthetic signal served as the actor network’s current state.Source: Data analysis · p. 32
• Canonical correlation analysis linearly aligned neural dynamics with two-dimensional behavior, then projected the rotationally derived signal into behavioral space to estimate its effect.The same linear transformations were used for the synthetic signal and neural-behavior alignment.Source: Data analysis · p. 32
• Individual-unit contribution scores combined loading onto the primary jPCA plane with activity coupling to population rotational phase, and units were ranked by score.Phase coupling was quantified by regressing unit activity on cosine and sine of the population phase.Source: Data analysis · p. 33
• Targeted pruning removed selected units’ input and output connections without retraining, after which task performance and rotational structure were reassessed.Performance was measured with task-specific metrics such as tracking error or reward, and jPCA was reapplied to recurrent activity.Source: Data analysis · pp. 33–34
• Preparatory and execution subspaces were constructed from condition-averaged activity using separate PCAs retaining components that explained 90% of variance, then compared through null-space variance and overlap.Overlap ranges from 0 for orthogonal subspaces to 1 for identical subspaces.Source: Data analysis · pp. 34–35
• A local linearization analysis estimated an empirical Jacobian by fitting a first-order linear dynamical system to condition-averaged hidden-state activity.Activity was segmented into preparatory and execution epochs based on task timing.Source: Data analysis · p. 35
• The analysis estimated separate Jacobians for preparatory and execution dynamics, examined their eigenspectra, and measured how neural modes and output gain related to output-potent and output-null subspaces.Local output sensitivity was quantified from the hidden-to-output Jacobian; output gain was computed from its projection into preparatory and execution subspaces.Source: Data analysis · pp. 35–36
• Neural and behavioral activity were reduced to low-dimensional manifolds, and canonical correlation analysis identified kinematics-aligned neural subspaces within preparatory and execution epochs.Cross-validated canonical correlations quantified alignment, while cross-epoch analyses tested its specificity.Source: Data analysis · p. 36
• Preparatory-to-execution neural displacement was measured as Euclidean centroid distances within trials and across the population, with significance assessed by permutation tests.The tests shuffled epoch labels within trials while preserving sample counts, then compared observed distances with a permutation null distribution.Source: Data analysis · pp. 36–37
• Neural activity was decomposed into rotational, output-potent, and condition-invariant components, whose behavioral predictive performance and geometric relations were evaluated using cross-validation and principal angles.The output-potent subspace came from SVD of a neural-to-target-error regression; the condition-invariant component was the leading principal component of across-condition mean activity.Source: Data analysis · p. 37
• Neural energy was calculated from latent activity of 512 RNN neurons, while muscle energy used W = I^2Rt to compare movement-strategy energy demands.The measures were used to examine how sensory-feedback configurations relate to neural-controller and muscle energy during movement.Source: Data analysis · pp. 37–38
• Rotational strength was quantified using smooth parametric curves to represent neural trajectories in neural state space, providing a consistent measure of rotation across different conditions.The angle ξ was computed between the first principal components of PCA subspaces; the remaining 55 time points, covering execution state, were treated as separate datasets.Source: Data analysis · p. 38