Communication in modular robotic motor control: Bilateral controllers under realistic constraints

Jingwen Li; Levin Kuhlmann; Jason Friedman; Gideon Kowadlo

cs.ROAug 14, 2026

Communication in modular robotic motor control: Bilateral controllers under realistic constraints

Authors: Jingwen Li, Levin Kuhlmann, Jason Friedman, Gideon Kowadlo

Organizations: Department of Data Science and AI, Monash University · Department of Physical Therapy & Sagol School of Neuroscience, Tel Aviv University · Cerenaut

Abstract

Robotic motor control in musculoskeletal systems requires fast, accurate movement and robust postural stabilization under signal-dependent noise (where motor command variance scales with command magnitude) and energetic cost. Modular controllers can distribute these competing demands across interacting submodules, but it remains unclear whether they outperform monolithic architectures under realistic constraints, and how inter-module communication shapes the resulting strategy. Inspired by the bilateral hemispheric organization of the brain, we introduce a recurrent controller of two GRU-based modules connected by a learnable, delayed inter-hemispheric channel, trained end-to-end in a differentiable two-arm musculoskeletal simulator. Across reaching and holding tasks, the modular architecture substantially outperforms a capacity-matched monolithic baseline. Compared to a matched modular controller without communication, learned inter-hemispheric communication reshapes the solution: improved endpoint precision, lower energetic cost in non-zero-delay regimes, and reduced muscle co-contraction. Our findings show that for robotics, biologically inspired modular controllers offer a practical route to robust movement under noise and energetic constraints, with inter-module communication providing a mechanism to tune trade-offs between precision, stability, and actuation cost.

Figures & tables

Figure 1: Musculoskeletal substrate and task design. (A) A MotorNet RigidTendonArm26 effector: a planar two-degree-of-freedom arm actuated by six Hill-type muscles—shoulder flexor/extensor (SF/SE), elbow flexor/extensor (EF/EE), and biarticular flexor/extensor (BF/BE). The bimanual setup places two such arms symmetrically in a shared planar workspace. (B) Center-out reaching task: the active hand moves from a central start position (green) to one of eight circularly arranged targets (red), while the resting hand remains undisturbed. (C) Postural holding task: the active hand maintains its position (green) against a zero-mean Gaussian external perturbation (red arrow) within a tolerance region (dashed circle), while the resting hand remains undisturbed. The tasks isolate trajectory- and stability-dominant demands, respectively.

Figure 2: Architecture overview of the bilateral controller. A shared sensory input is broadcast to two GRU-based hemispheric controllers linked by delayed inter-module communication. Their outputs are combined into a 12-dimensional muscle-command vector, passed through signal-dependent multiplicative motor noise, and split into six muscle activations for each arm. The diagram highlights the three constraints used in this study: delayed inter-module communication, signal-dependent motor noise, and energetic cost in the training objective.

Figure 3: Task-performance profiles across controller variants and constraint settings. (A) Distribution of success rates across trained configurations from the shared noise–energy grid and ten random seeds. Boxes show medians and interquartile ranges, whiskers show non-outlier spread, and points show individual trained configurations. Bilateral w/ CC is split into zero-delay ( $d=0$ ) and non-zero-delay ( $d>0$ ) training conditions, while Bilateral w/o CC and Monolithic denote the communication-free modular and monolithic baselines. (B) Constraint-dependent success rate landscapes across energetic cost and motor-noise settings. Color and overlaid values report mean success rate across ten seeds for each parameter combination. The monolithic controller shows its highest fragility at intermediate noise, whereas modular controllers maintain higher success rate across the same regimes.

Metric	Bi w/ CC, $d>0$	Bi w/ CC, $d=0$	Bi w/o CC	Mono.
Success rate $\uparrow$	88.81% $\pm$ 10.01%	77.53% ∗ $\pm$ 21.99%	87.59% ∗ $\pm$ 13.82%	68.91% ∗ $\pm$ 27.81%
Endpoint error $\downarrow$	0.0085 $\pm$ 0.0017	0.0091 ∗ $\pm$ 0.0022	0.0117 ∗ $\pm$ 0.0041	0.0091 ∗ $\pm$ 0.0023
Trajectory error $\downarrow$	0.0022 $\pm$ 0.0003	0.0025 ∗ $\pm$ 0.0005	0.0046 $\pm$ 0.0025	0.0026 ∗ $\pm$ 0.0005
Move. variance $\downarrow$	0.0081 $\pm$ 0.0013	0.0092 ∗ $\pm$ 0.0038	0.0093 $\pm$ 0.0027	0.0099 ∗ $\pm$ 0.0030
Hold variance $\downarrow$	1.054e-5 $\pm$ 6.723e-6	1.556e-5 ∗ $\pm$ 1.471e-5	1.128e-5 $\pm$ 8.132e-6	1.867e-5 ∗ $\pm$ 1.338e-5
Energetic cost $\downarrow$	1.0875 $\pm$ 1.6950	1.3196 ∗ $\pm$ 1.7233	1.1930 $\pm$ 1.8288	1.7666 ∗ $\pm$ 1.9085

Table 1: Metric decomposition across matched settings (ten seeds, four

\lambda

values, four noise levels;

N=160

). Bilateral w/ CC (

d>0

) averages each metric over

d\in\{5,10,20\}

per matched setting before paired testing and is the reference column for all asterisks. Asterisks mark scores in the other columns that differ significantly from Bilateral w/ CC (

d>0

) under Bonferroni-corrected paired Wilcoxon tests (

p<0.05

, corrected over the three column-wise comparisons within each metric). Arrows indicate the direction of improvement (

\uparrow

higher is better,

\downarrow

lower is better), and bold marks the best mean in each row.

Figure 4: Energetic cost, CCI, and success rate across constraint regimes. Panels within each row share a common y-axis scale. Each column corresponds to a noise level ( $k=0,0.05,0.1,0.2$ ); within each panel, bars show the four model conditions (Bilateral w/ CC averaged over $\delta\in\{5,10,20\}$ , Bilateral w/ CC at $\delta=0$ , Bilateral w/o CC, Monolithic) at each energy-penalty level. Non-zero-delay Bilateral w/ CC shows its clearest gains at low-to-moderate noise; at high noise, success rate becomes similar across modular conditions but Bilateral w/ CC variants retain more active actuation.

Figure 5: Generalization results for success rate on unseen reach and hold conditions. Filled bars: out-of-distribution performance; faint bars: matched in-distribution baseline. Bar labels report average success rates. For Bilateral w/ CC, zero-delay models ( $d=0$ ) are separated from delayed models ( $d>0$ , averaged over $d\in\{5,10,20\}$ ). On reach-16 , success uses a fixed reach threshold of $r20=20$ mm rather than a radius-scaled threshold. On hold-2N , the Gaussian holding perturbation has per-axis standard deviation 2 N. In both conditions, non-zero-delay Bilateral w/ CC performs best, followed in order by Bilateral w/o CC, zero-delay Bilateral w/ CC, and Monolithic.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Success decomposition for the monolithic baseline under the original Gaussian hold perturbation with per-axis standard deviation $0.5\,\mathrm{N}$ . The high-noise rebound in success rate is driven mainly by hold success, while reach success remains high across noise levels.

Figure 7: Perturbation-strength diagnostic for the monolithic baseline. Left: with Gaussian external force standard deviation increased to $1.5\,\mathrm{N}$ per axis, the high-noise rebound remains visible. Right: with standard deviation increased to $4.0\,\mathrm{N}$ per axis, the same high-noise region collapses toward floor performance.

Figure 8: Energy–variance diagnostic for the high-noise regime. High-noise solutions cluster near low energetic cost and low hold variance, consistent with a low-output holding regime rather than robust active stabilization.

Robotic Control

Robotic control studies how machines use sensing and computation to produce reliable movement and interaction in changing environments.

See all 584 papers

Oct 24, 2026

Learning Functional Specialisation for Bimanual Robotic Control

Bimanual robotic control requires simultaneous management of distinct functional demands under motor noise, energetic costs, and communication delays. We investigate how learning responsibility should be organised across modular controllers and the two arms in a simulated musculoskeletal system with two communicating recurrent controllers. Using a 2 x 2 design, we compare differential versus shared controller objectives and fixed versus balanced hand-task assignments. Task-aligned functional specialisation was strongest when differential objectives were combined with a fixed assignment aligned with contralateral control routing. This condition achieved the highest canonical-task performance and control efficiency, reaching 81.33% bimanual success. Its advantage over shared objectives increased with motor noise but narrowed under stronger energy penalties. Stronger specialisation also reduced hand-role flexibility: reversing the hand roles decreased bimanual success by 80.46 percentage points under Differential-Fixed training, compared with 1.51 percentage points under Shared-Balanced training. These results show that controller objectives and hand-task experience jointly determine functional organisation. Stronger role-specific specialisation favours performance and efficiency when roles remain stable, whereas more distributed training favours interchangeability when roles change.

Jingwen Li, Levin Kuhlmann, Jason Friedman +1

Department of Data Science and AI, Monash University · Department of Physical Therapy & Sagol School of Neuroscience, Tel Aviv University · Cerenaut

Robotic Control Bimanual Robotic Manipulation

Ask PDF

Apr 23, 2026

Insect-inspired modular architectures as inductive biases for reinforcement learning

Most reinforcement-learning (RL) controllers used in continuous control are architecturally centralized: observations are compressed into a single latent state from which both value estimates and actions are produced. Biological control systems are often organized differently. Insects, in particular, coordinate navigation, heading stabilization, memory, and context-dependent action selection through distributed circuits rather than a single monolithic controller. Motivated by this contrast, we study an RL policy architecture that decomposes control into interacting modules for sensory encoding, heading representation, sparse associative memory, recurrent command generation, and local motor control, with a learned arbitration mechanism that allocates motor authority across modules. The model is evaluated on a two-dimensional navigation task that require simultaneous food seeking, obstacle avoidance, and predator escape. In a six-seed predator-navigation experiment trained with Proximal Policy Optimization (PPO) for 75 updates, the modular policy achieves the strongest final mean performance among the tested controllers, with final episodic return

-2798.8\pm964.4

versus

-3778.0\pm628.1

for a centralized gated recurrent unit (GRU) and

-4727.5\pm772.5

for a centralized multilayer perceptron (MLP). The modular policy also attains the lowest final value loss and stable PPO optimization statistics while driving module-assignment entropy to

0.0457\pm0.0244

, indicating highly selective control allocation. These results suggest that distributed control can serve as a useful inductive bias for RL problems involving dynamically competing behavioral objectives.

Anne E. Staples

Department of Mechanical Engineering, Virginia Tech, Blacksburg, Virginia 24061, USA

RL Control Reinforcement Learning Policy Learning

Ask PDF

Oct 7, 2026

Decoding Neural Population Dynamics through Robotic Analog

Animal evidence shows that precise voluntary movements arise from rotational neural population dynamics in motor cortex, but their physical effects remain unknown. We developed a robotic analog of biological motor systems with artificial muscles, multimodal sensors, and a neural network controller trained via reinforcement learning. The robotic analog exhibited accurate movements, robustness to damage, and neural population dynamics akin to animals. This task-driven, embodied model illuminates the causal link between neural population dynamics and motor outcomes. We discovered that neural rotations generate oscillatory maneuvers orthogonal to the reaching direction, optimizing trajectory adjustments, which is confirmed by primate neural data. The model also revealed counterintuitive neural energy principles under sensor and motor redundancies, and striking Eureka moments during motor learning, bridging biological and artificial systems. These findings provide new perspectives on how neural dynamics contribute to accurate and flexible movement, inspiring future intelligent robots with animal-like mobility.

Wenhui Chen, Jiyue Tao, Yitao Cheng +4

School of Advanced Manufacturing and Robotics, Peking University, Beijing & 100871, China. · School of Life Sciences, Peking University, Beijing & 100871, China. · IDG/McGovern Institute for Brain Research, Peking University, Beijing & 100871, China. +2

Dynamical Systems

Ask PDF

Abstract

Robotic Control

Explore similar work

Learning Functional Specialisation for Bimanual Robotic Control

Insect-inspired modular architectures as inductive biases for reinforcement learning

Decoding Neural Population Dynamics through Robotic Analog