acceptodds
Withdrawn from ICLR 2027

Superfly: Learning Locomotion with a Fly Connectome

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We present Superfly, a sensorimotor network built from the adult Drosophila brain and ventral nerve cord connectome for embodied locomotion. The policy maps inputs to body controls through about 23,000 neurons and 2.2 million directed edges, fixing each connection's endpoints, synapse count, and excitatory or inhibitory sign. Training learns nonnegative scales that convert fixed synapse counts into synaptic currents. Mappings based on known sensorimotor pathways route body feedback, movement commands, and previous controls to corresponding neuron populations. Each neuron maintains voltage across body steps, and decoder readouts map its activity to controls for a MuJoCo fly model. Behavioral cloning learns about 37,000 parameters from demonstrations by a flybody walking expert. Compared with a parameter-matched graph neural network (GNN), the policy achieves lower validation action error (MSE 0.069 versus 0.096) and travels farther in five-second walking trials (8.5 versus 0.4 cm toward a 10 cm target). Although training provides only expert body controls and no neural activity targets, turn commands selectively modulate neuron types with established roles in navigation and steering. These responses persist after rewiring, showing that turn-command selectivity does not require the exact synaptic partners. The rewired policy fits actions better offline but travels less in closed loop. This executable model links learned locomotion to identified biological neurons and predicts their activity during behavior. We make our code, checkpoints, and an interactive website for exploring neuron activity available online at https://superfly-brain.vercel.app.

Primary area:
applications to neuroscience & cognitive science
Keywords:
connectome-constrained learning, Drosophila, recurrent neural networks, behavioral cloning, embodied locomotion, computational neuroscience

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jrjVOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors take a subset of the full wiring diagram of the fruit fly, Drosophila melanogaster, and use it as the architecture of a controller that is trained by behavioral cloning to drive a biomechanical fly model in MuJoCo. The connectome fixes which neurons connect and with what sign; training fits input routes, cell-type physiology, synaptic scales, and motor readouts. The resulting policy walks, and the authors read out activity in known steering neurons as evidence that the model is biologically interpretable.

Critical strengths
  1. The model is tied to identified neurons, so in principle its activity can be compared against experiments.
  2. The matched-parameter baselines and the degree-preserving rewired control are a reasonable experimental design.
  3. The paper is easy to read.
Critical weaknesses
  1. The model keeps 22,688 neurons, whereas the fly brain (e.g., the FAFB dataset) alone has about 140k neurons. The paper does not explain why these seven populations were picked and what was dropped, nor how much of the connectome the retained subset actually covers.
  2. It takes a trip to the appendix (Fig. 5, App. B.1) to work out what the 75 "body readings" are, that the 59 controls are joint/adhesion/head-abdomen targets rather than muscles, and what each of the seven linear readouts does. These should be spelled out up front where the architecture is introduced.
  3. Initialization, learning rates, and clamping are described, but only one seed is ever trained. Task-optimized connectome models can settle into different solutions because the fly brain carries many redundant, parallel pathways. Lappalainen et al. [1] and Ozdil et al. [2], for example, trained an ensemble of models for exactly this reason. Drawing conclusions about specific cell types in Fig. 4 from one network is therefore problematic.
  4. The yaw command is fed only into FC2, and the analysis then checks whether cell types known to sit downstream of FC2 respond. That is close to guaranteed by the anatomy. What is missing is a check that descending neurons which should not carry turning signals, for example MDN [3] or the antennal-grooming aDNs [4], stay quiet. Comparing against a population mean is not a control.
  5. Only proprioceptive and inertial feedback is closed through MuJoCo. There is no exteroceptive input (vision, olfaction), so the compass and goal circuits the interpretation leans on never receive the kind of sensory drive they get in the animal.
  6. What do we exactly learn from this study? Embodied connectome demonstrations are no longer rare and a paper submitted to ICLR should say what new insight we gain from such work. No prediction about the biological circuit is made and tested; the MLP baseline outperforms the connectome model on every metric; and the rewired network keeps the same cell-type selectivity. By contrast, recent mechanistic work on this same circuit, such as the network model of path control by Osman et al. [5], is built around experimentally testable mechanisms. The present paper does not reach that standard.

References

[1] Lappalainen, J.K., Tschopp, F.D., Prakhya, S., et al. (2024). Connectome-constrained networks predict neural activity across the fly visual system. Nature 634.

[2] Ozdil, P.G., Arreguit, J., Scherrer, C., Ijspeert, A.J., Ramdya, P. (2026). Centralized brain networks underlie body part coordination during grooming. Nature Communications 17. doi: 10.1038/s41467-026-72152-x.

[3] Bidaye, S.S., Machacek, C., Wu, Y., Dickson, B.J. (2014). Neuronal control of Drosophila walking direction. Science 344.

[4] Hampel, S., Franconville, R., Simpson, J.H., Seeds, A.M. (2015). A neural command circuit for grooming movement control. eLife 4, e08758.

[5] Osman, M.A.M., Zhang, W., Pehlevan, C., Wilson, R.I. (2026). A network model of path control. bioRxiv, doi:10.64898/2026.10.02.756268.

Questions for authors
  1. Why were these seven populations kept and the rest of the brain discarded? What fraction of BANC does the model represent?
  2. Can the authors train several seeds and report the variability of the Fig. 4 cell-type responses across them?
  3. Do non-steering descending neurons (e.g., MDN, aDN) respond to the yaw intervention?
  4. Why was no exteroceptive feedback included, given that the simulators cited already support it?
  5. What specific, testable prediction about the fly’s steering circuit does this model make that was not already known from the anatomy?
Reviewer_x6wiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SUPERFLY is a walking policy for the flybody MuJoCo fly, trained by cloning its walking expert. It uses a fixed BANC subgraph (22,688 neurons, 2.2M signed edges), with one passive scalar-voltage state per neuron. Its 37k learned parameters cover group-shared synaptic scales, cell-type physiology, anatomically routed inputs, and descending/motor readouts. Velocity commands enter FC2; previous controls enter the readout populations.

The baselines are a parameter-matched MLP, a same-graph GNN, and a degree-preserving rewired Null. SUPERFLY walks in closed loop while the GNN does not, and travels farther than Null despite higher offline MSE. The MLP is strongest overall. Yaw commands preferentially recruit PFL2/3 and DNa02/03, but this enrichment survives rewiring.

Critical strengths
  1. Identifiable internal states in a working controller. Each state belongs to a specific BANC neuron, giving recordings and targeted perturbations a direct counterpart in the model. The main differences from FlyGM are scalar point-neuron dynamics, nerve-cord coverage, and anatomical input routing.
  2. Fair comparisons and candid reporting. Parameter count, data, training budget, and checkpoint selection are matched. I appreciate the clear reporting of the stronger MLP and Null’s lower offline MSE. The latter, alongside worse closed-loop travel, is a useful result.
Critical weaknesses
  1. The contribution of exact wiring remains unclear. Previous controls enter the same populations used for readout, creating an action-history pathway that need not cross any inter-neuronal connection. How much of the action fit comes from this pathway?

    There is also only one trained instance per policy and one rewire. Figure 3 varies initial posture, not training seeds or rewires. The authors acknowledge this, but I am not yet convinced that the native–Null gap establishes a reproducible benefit of exact connectivity. Coarse anatomy, input routing, and action history remain difficult to separate.

  2. The yaw responses are not yet convincing biological predictions. The analysis uses unsigned current contrasts, commands of ±1 rad/s versus a trained range of ±0.35 rad/s, and only five states. Enrichment also survives rewiring. I would want to see whether response signs match biology and remain consistent across states. This is a central issue when interpretability is the main advantage over the stronger MLP.

Questions for authors
  1. Action history. What validation MSE does copying the last action achieve? How do MSE and closed-loop walking change after retraining without the efference copy?
  2. Seeds and rewires. Can you compare native and Null across matched training seeds and independent rewires, e.g., 3–5 each? How many edges were swapped, what fraction of original edges remains, and what are Null’s yaw-response values for PFL2/3 and DNa02/03?
  3. Signed responses. Across more states and commands within ±0.35 rad/s, what are the left–right differences in release changes for DNa02 and PFL3 grouped by LAL output side? Are the signs consistent across states and with experimentally reported turn direction?
  4. GNN and evaluation. How are the GNN’s 37k parameters allocated and shared, and how sensitive are results to its configuration? How do the 25 held-out postures for the 10-s evaluation relate to the 20 + 5 demonstration postures? Are both Table 1 MSEs recomputed at the selected checkpoint using the same procedure?
  5. Beyond FlyGM. What new capability is demonstrated by scalar dynamics, nerve-cord coverage, and anatomical routing?
Reviewer_FzxcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary
  • This paper present Superfly, a network built from the BANC fly connectome and trained to walk using flybody, the Janelia/DeepMind fly body model in MuJoCo.
  • Network architecture:
    • Body sensor readings (75 dimensions) are mapped by learned linear maps onto anatomically matched sensory neurons, which are selected using the connectome’s organ, leg, and side annotations. Each input channel delivers the same current to all of its recipient neurons.
    • The velocity command (3 dimensions) is input to the FC2 neurons of the fan-shaped body, also via a learned linear map.
    • The previous action (59 dimensions) is fed back as input to the descending and leg motor neurons.
    • Only a small subset of the connectome is modeled (~15% of BANC). It comprises the central complex plus a few relay neurons, all descending and ascending neurons, leg motor neurons, a subset of VNC interneurons, and leg, haltere, and antennal mechanosensory neurons; the rest of the brain is excluded. Connectivity, synapse counts, and transmitter-derived signs are fixed. Each neuron is a non-spiking leaky point neuron with a scalar voltage. Training fits only per-cell-type resting voltages and time constants, plus non-negative synaptic scales shared within groups defined by source cell type and target population.
    • Linear readouts map each leg’s motor neurons to that leg’s joint and adhesion controls, and map all descending neurons to the head and abdomen controls.
    • The network is trained by behavioral cloning of a pretrained flybody walking policy, rather than with the RL-based imitation learning used in related work.
    • Performance is evaluated by the mean squared error of one-step action prediction on held-out starting poses, and by the distance traveled in closed-loop rollouts given a straight-walking command.
  • Superfly is compared to several parameter-matched baselines:
    • An MLP. Its architecture is not reported; the parameter count is consistent with a single hidden layer.
    • A GNN on the same connectome graph, in which each neuron carries a learned feature vector (~50 dimensions) updated by message passing, instead of a single scalar voltage.
    • A null model created by repeatedly swapping the postsynaptic targets of pairs of connections within the same population block: a->b and c->d, with weights w_ab and w_cd, become a->d and c->b, with weights w_ab and w_cd. This preserves every neuron’s in- and out-degree.
    • The MLP performs best on every metric, and the GNN performs poorly. The null model fits actions slightly better than Superfly offline but drifts more in closed loop, walking in circles.
  • The authors examine how a yaw command propagates through the network and find that known steering neurons (PFL2/3, DNa02/03) are among the most responsive. This is consistent with their known roles.
Critical strengths
  • The goal is interesting and timely: the paper builds a model constrained by both the brain and the VNC connectome and connects it to a physics-simulated body for closed-loop locomotion.
  • The number of learned parameters is kept to a minimum given current knowledge. Wiring, synapse counts, and signs are fixed from the connectome, and only quantities the connectome cannot provide (per-cell-type physiology, shared synaptic scales, and the input and output maps) are fit.
  • The controls and reporting are honest and useful. The authors include a parameter-matched MLP and report that it outperforms their model, use a degree-preserving null model, separate offline imitation error from closed-loop behavior, and state their limitations clearly.
Critical weaknesses
  • The parameter-matched MLP outperforms Superfly on every metric. This is an interesting negative result, but the paper does not present it as one; it still presents Superfly as a useful approach, without evidence that the connectome constraint improves control.
  • The steering-neuron analysis is largely determined by construction. Because most of the brain is not modeled and the command is injected directly at FC2, the signal can only reach the motor output through the known steering pathway, so it is unsurprising that these neurons respond.
  • Details are missing about which neurons of the central nervous system were selected for modeling and how they were selected. The central complex is known to support navigation (for example, tracking heading and goal direction), not the moment-to-moment control of walking. It is also not the only brain region involved in these behaviors: there are indirect connections between the central complex and the descending neurons that are likely important, and these are mostly excluded from the model. More generally, we do not know enough about the fly brain to conclude that the modeled regions are the only relevant ones.
  • Only one null model is evaluated. Given that it performs reasonably well, it would be useful to compare multiple rewired networks, and multiple training seeds, to determine whether the differences from the native connectome are meaningful.
  • The GNN baseline appears weak and needs a more complete description, including its architecture, how parameters are shared across neurons, and how its hyperparameters were chosen.
  • The evaluation metric, distance traveled, is not the only reasonable choice. Because the policy is commanded to walk at a fixed velocity, velocity tracking error (for example, averaged over windows of about one stride) would be more appropriate.
  • Routing the previous action directly into the descending and motor neurons that are read out provides a shortcut that bypasses the rest of the modeled nervous system.
  • The authors should explain why they use behavioral cloning rather than reinforcement learning. The closed-loop failures they observe are characteristic of behavioral cloning.
  • The released Janelia/DeepMind flybody policy checkpoints were trained with DMPO, not SAC, as the paper states.
  • Missing details include the MLP architecture and a full specification of the learned input and output maps.
Questions for authors
  • How were the modeled neurons selected, and which neurons were included? For each type of retained neuron, what fraction of its input and output synapses were lost because the partner neurons were not modeled?
  • What is the architecture of the learned maps from the inputs to the connectome and from the connectome to the outputs?
  • What is the architecture of the MLP baseline?
  • What is the architecture of the GNN baseline?
  • How important is the direct input of the previous action to the descending and motor neurons? How does performance change when this input is removed?
  • How many swaps were performed to create the null model, and what fraction of connections were rewired? I- n the yaw analysis, why was a yaw command of +/-1 rad/s used, when the largest value used in training was 0.35 rad/s? How many neurons can be reached from FC2 within three synaptic steps?
  • Why was the model trained with behavioral cloning rather than with reinforcement-learning-based imitation learning, as used to train the flybody teacher policy and FlyGM?
  • It's worth discussing the results wrt the Digital Sphinx paper by Brunton et al. 2026. Superfly’s null and MLP results are consistent with this argument, and the authors should address it directly.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.