PaperlayerView exact revision on arXiv

Source-linked AI summary

FlyCNS: Connectome-Grounded Information Organization for Communication-Constrained Embodied Control

Jinchang Zhang, Jiakai Lin, Guoyu Lu

arXiv:2609.28816v1cs.RO

TL;DR

Learning-based control often centralizes information despite robots’ distributed bodies, leaving open which information must cross local-module boundaries for coordination. FlyCNS uses a connectome-derived directional prior to organize learned ascending and descending communication; in Go1 simulation, it maintains tracking with reduced communication under the tightest budget.

  • Problem

    Distributed robot bodies need local sensorimotor responses and cross-limb coordination, but centralized policies expose complete body state and do not determine which information must cross module boundaries.

    Source: §I · p. 1

  • Method

    FlyCNS preserves limb-local computation and uses a directional prior from the Drosophila brain–cord connectome to bias learned ascending and descending communication.

    Source: §I · p. 1

  • Results

    At the most restrictive budget, FlyCNS uses approximately 21–22% of FULL communication while maintaining a tracking score of about 0.882 under both command protocols, within 6.1% of FULL.

    Source: §I · p. 2

  • Takeaways & Limitations

    The connectome supplies a transferable information-organization prior, while message content, transmission timing, and motor behavior remain task-adaptive.

    Source: §I · p. 1

Abstract

from arXiv · show

Robotic bodies are inherently distributed in sensing and actuation, yet learning-based control still commonly relies on centralized information processing. This work studies the problem of information organization in communication-constrained embodied control: which computations should remain local, and which information is worth transmitting for whole-body coordination. We propose FlyCNS, an embodied information-organization framework inspired by the Drosophila brain--nerve-cord connectome. FlyCNS preserves local sensorimotor computation within each limb and enables selective long-range communication through separate ascending and descending routing pathways. From a real connectome, FlyCNS extracts the directional structural complexity of these two pathway types and uses it as a weak prior over communication allocation, while message content, transmission timing, and locomotion policies remain task-adaptive and are learned through reinforcement learning. In Unitree Go1 simulation, FlyCNS exhibits more graceful performance degradation as the communication budget is tightened. Under the most restrictive setting, it uses only about 21--22\% of the communication of the full-communication reference, while still maintaining a tracking score of approximately 0.882 under both command protocols, with a gap of no more than 6.1\% from the full-communication reference. These results indicate that real neural connectomes can inform not only the structural design of control networks, but also provide transferable inductive biases for information organization across embodiments, guiding robots in balancing local computation and long-range coordination under limited communication resources.

I. INTRODUCTION

FlyCNS treats communication-constrained control as an information-organization problem: what computation stays local and what information crosses modules for coordination. It uses a directional connectome-derived prior to organize learned communication, and maintains locomotion tracking with substantially reduced communication in simulation.

I. INTRODUCTION

  • Centralized policies expose complete body state to one controller, while distributed control raises the question of which information must cross module boundaries for effective coordination.The Go1 communication budget models this bottleneck; the paper does not assume its stock bus is saturated.Source: §I · p. 1
  • FlyCNS uses directional structural statistics from the Drosophila brain–cord connectome as a weak prior without transferring individual neurons or biological circuits.Message content, routing decisions, and motor control remain task-adaptive and are learned through robot–environment interaction.Source: §I · pp. 1–2
  • FlyCNS preserves limb-local sensorimotor computation and uses selective ascending feedback and descending coordination through a global coordinator.The two communication pathways share a budget, while their transmission and the motor policy are learned.Source: §I · pp. 1–2
  • At the most restrictive budget, FlyCNS uses approximately 21–22% of FULL communication while maintaining a tracking score of about 0.882 under both command protocols, within 6.1% of FULL.It also exhibits more graceful degradation as bandwidth is reduced.Source: §I · p. 2
  • The paper frames communication-constrained embodied control as allocating information between local sensorimotor computation and long-range coordination.This makes the local–global information boundary an explicit part of the control policy.Source: §I · pp. 1–2

II. RELATED WORK

Prior work studies connectomic control and modular communication, but leaves open how neural brain–cord organization can become a transferable routing prior. FlyCNS abstracts directional information flow into local computation, learned routing, and connectome-biased allocation under a communication budget.

A. Biological/Connectomic Control

  • Connectome-based controllers such as FlyGM and FLYNN use neural topology, but leave open how brain–cord organization can yield a transferable routing prior for robots with different morphology.Biological and neuromechanical work also links wiring to sensorimotor computation and realistic bodies.Source: A. Biological/Connectomic Control · p. 2

B. Modular/Decentralized Robot Learning

  • Modular policies use shared parameters, local controllers, and learned or selective messages, while FlyCNS adds a connectome-derived directional prior for budgeted communication.Prior work includes morphology transfer, local-information control, and bandwidth-aware event-triggered communication.Source: B. Modular/Decentralized Robot Learning · p. 2

III. METHOD

  • FlyCNS transfers a coarse directional organization principle—not morphology or biological circuitry—while message semantics, timing, and motor behavior remain task-adaptive.Its three stages are limb-local computation, NeuroRoute's learned boundary crossing, and connectome-biased allocation of ascending and descending traffic.Source: §III · p. 2

A. From Neural Organization to Embodied Information Architecture

  • The architecture maps local sensorimotor circuits to limb modules, ascending pathways to body-to-global flow, and descending pathways to global-to-body coordination.In Go1, four limb modules control three actuated joints each; remote information crosses an explicit local–global boundary under an average communication budget.Source: A. From Neural Organization to… · p. 3

B. Local Sensorimotor Autonomy and Global Coordination

Each limb encodes local observation with its identity to form an ascending message, while a global coordinator integrates body-level observations and available limb messages to generate descending context. These two message transitions augment rather than replace local sensorimotor computation.

B. Local Sensorimotor Autonomy and Global Coordination

C. NeuroRoute: Selective Bidirectional Coordination

NeuroRoute learns independent ascending and descending transmission decisions that follow the control loop’s causal order. It retains cached messages when transmission is skipped and uses a training curriculum to stabilize learning as communication becomes sparse.

C. NeuroRoute: Selective Bidirectional Coordination

  • Ascending and descending routers make independent decisions because their messages update body knowledge and local-action context at different points in the control loop.Source: C. NeuroRoute: Selective Bidirectional… · p. 3
  • Ascending gates are sampled first, and the resulting cache state informs global representation and conditional descending routing probabilities.When a message is not transmitted, the receiver retains its most recent value; message age provides temporal context.Source: C. NeuroRoute: Selective Bidirectional… · p. 4
  • Training progressively relaxes the minimum transmission probability from full communication toward the learned policy as the budget tightens.The curriculum stabilizes learning under severe sparsity without changing the final routing objective.Source: C. NeuroRoute: Selective Bidirectional… · p. 4

D. Connectome-Grounded Directional Prior

FlyCNS derives separate ascending and descending structural statistics from the Drosophila brain–cord connectome and uses them as a weak prior on communication allocation. The prior biases the directional distribution of communication, while message content and timing remain task-adaptive.

D. Connectome-Grounded Directional Prior

  • The connectome supplies a task-independent directional preference before robot optimization, rather than a robot communication policy learned from the connectome itself.Source: D. Connectome-Grounded Directional Prior · p. 4
  • The directional representations summarize connectivity from three leg-related sensory groups to 16 ascending super-clusters and from 16 descending super-clusters to three leg-related motor groups.Both matrices are extracted and normalized using the same frozen type-balanced protocol before robot training.Source: D. Connectome-Grounded Directional Prior · p. 4
  • The connectome prior biases how scarce long-range communication is allocated between ascending and descending pathways, without fixing individual transmissions.Source: D. Connectome-Grounded Directional Prior · p. 4
  • The objective separates a total-communication constraint from a biological term that biases directional organization.Message content and timing remain state-dependent and task-adaptive.Source: D. Connectome-Grounded Directional Prior · p. 4

A. Experimental Protocol

The protocol evaluates Go1 locomotion in MuJoCo Playground under multiple command modes and communication configurations. It reports tracking and motion-quality measures alongside realized communication, with results summarized across training seeds.

A. Experimental Protocol

  • The Go1JoystickFlatTerrain task uses a 12-joint-target policy trained with PPO, a privileged critic, and running observation normalization.Each training run consists of 206,438,400 environment interactions.Source: A. Experimental Protocol · p. 4
  • The comparisons include a centralized actor, an all-send modular reference, a learned bidirectional router, and a calibrated periodic schedule.FULL transmits all ascending and descending messages at every control step; PERIODIC’s schedule is calibrated from ASYM communication records, not task performance.Source: A. Experimental Protocol · p. 4
  • Eight fixed resets per command mode are evaluated at each checkpoint, and metrics are averaged across resets and training seeds.Rollouts last 1,000 control steps; the scripted mode switches commands every five seconds, and error bars use sample standard deviation across three seeds.Source: A. Experimental Protocol · pp. 4–5
  • The protocol measures tracking, realized communication, motion feasibility, and directional traffic across communication budgets and Official/Scripted commands.Table II identifies Rv as realized-to-commanded planar speed, NS as near-static percentage, ev and eω as velocity and yaw-rate RMSE, and C↑/C↓ as ascending and descending communication.Source: A. Experimental Protocol · p. 5, Table II · p. 6

B. Performance–Communication Trade-off

FlyCNS improves tracking while reducing measured communication relative to baselines, with the clearest joint advantage at 50% and 25% budgets. Its results also show that a small tracking disadvantage can occur in an individual 75% comparison.

B. Performance–Communication Trade-off

  • At 50% and 25% budgets, FlyCNS achieves higher tracking performance and lower communication cost than either baseline across all seed–mode comparisons.At 75%, one Scripted comparison with PERIODIC has a 0.000123 lower FlyCNS tracking score, though FlyCNS still uses less communication.Source: B. Performance–Communication Trade-off · p. 6
  • Against PERIODIC, Official-command tracking improvements rise from 0.0095 at 75% to 0.0988 at 50% and 0.3344 at 25%.Communication is reduced by 11.17%, 6.59%, and 12.81%, respectively.Source: B. Performance–Communication Trade-off · p. 6
  • Against PERIODIC under Scripted commands, tracking improves by 0.0054, 0.0278, and 0.0421 at 75%, 50%, and 25%, respectively.Communication is reduced by 14.46%, 11.69%, and 16.04%, respectively.Source: B. Performance–Communication Trade-off · p. 6

C. Control Quality under Increasing Communication Scarcity

As communication becomes scarce, FlyCNS maintains feasible locomotion more consistently than ASYM and PERIODIC, while its own tracking performance declines modestly with tighter budgets. Rollouts show that low-bandwidth failures can appear as command-tracking drift rather than falls.

C. Control Quality under Increasing Communication Scarcity

  • At 25%, FlyCNS records Rv = 0.916/0.912 and near-static fractions of 0.019%/0.128% under Official/Scripted commands, using 22.16%/21.34% of FULL communication.All three FlyCNS seeds meet the numerical locomotion criteria in both modes; ASYM fails in both, while PERIODIC passes for only one Scripted seed.Source: C. Control Quality under Increasing… · p. 6
  • During lateral, yaw, and combined-motion phases, ASYM and PERIODIC trajectories increasingly diverge from command-integrated references, while FlyCNS stays closer.Low-bandwidth failure may leave the robot upright and moving while it no longer faithfully executes the command.Source: C. Control Quality under Increasing… · p. 7, Figure 3 · p. 7
  • From 50% to 25%, FlyCNS reduces realized communication by 52.91%/52.04% under Official/Scripted commands, while tracking scores drop by only 0.0367/0.0377.FlyCNS does not eliminate the communication–control trade-off, but degrades more gracefully as communication becomes scarce.Source: C. Control Quality under Increasing… · p. 7

D. Learning Dynamics, Traffic, and Attribution

With every policy evaluated at the final 25% communication target, FlyCNS improves sharply late in training and finishes above PERIODIC and ASYM on Official tracking.

D. Learning Dynamics, Traffic, and Attribution

  • At the final 25% target, FlyCNS Official tracking rises from 0.0834 at 91.75M interactions to 0.8815, above PERIODIC at 0.5471 and ASYM at 0.4359.Intermediate FlyCNS scores are 0.5452 at 114.69M and 0.8569 at 160.56M interactions; a similar late-stage improvement occurs under Scripted commands.Source: D. Learning Dynamics, Traffic, and… · p. 7

V. CONCLUSION

The conclusion frames communication-constrained embodied control as deciding what computation stays local and what information should be transmitted for coordination. FlyCNS uses a connectome-derived directional prior while learning routing and control from robot interaction.

V. CONCLUSION

  • FlyCNS combines limb-local sensorimotor computation with selective ascending and descending communication, using a directional prior extracted from the Drosophila brain–cord connectome.It does not copy biological circuitry: message semantics, transmission timing, and motor behavior remain learned from robot interaction.Source: §V · p. 8
  • On Unitree Go1, FlyCNS improves the performance–communication trade-off as communication becomes scarce, retaining effective locomotion with about 21–22% of full-communication traffic at the most restrictive setting.The conclusion presents measured neural information-organization principles as transferable inductive biases for distributing information across robot bodies.Source: §V · p. 8
Loading 2609.28816v1…