Fly Whiskers Meet LiDAR: Robust Navigation under Sensor Failure

arXiv:2611.03897 · q-bio.NC · Submitted 2026-11-02 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: I'm Ines.

Marcus: And I'm Marcus, here with Yuki.

Ines: We're looking at the paper: "Fly Whiskers Meet LiDAR".

Marcus: Here's the gist of the paper. Fly Whiskers Meet LiDAR: Robust Navigation under Sensor Failure extends FLYNN by integrating a 32-ray 2D LiDAR interface into a fly connectome network to improve navigation when vision fails The strongest benefits occur under total vision loss, out of distribution (OOD) textures, and periodic dropout FLYNN-LiDAR achieves 96.6% success under full vision, whereas the baselines reach 44.6% to 56.6% with the scan on The result is not universal superiority, as MobileNet retains a significant left-eye-off advantage Sensor Integration and Mechanism The LiDAR interface involves injecting a fixed, spatially ordered projection of a 32-ray 2D LiDAR scan into the 37 BM Vib head-bristle neurons This projection has fixed weights and one learned scalar input gain, which is integrated in the recurrent network The scan is represented by encoding each continuous range as 1 − r/8, where zero denotes free space and one denotes contact Rays are assigned by hemifield to the nearest of 18 left and 19 right BM Vib afferents in a fixed midline-to-lateral map The LiDAR-to-whisker mapping preserves true anatomical laterality, unlike the small-world control where laterality is assigned by convention Performance Under Stress Conditions The robustness of FLYNN-LiDAR is demonstrated across several failure modes and conditions The strongest benefits occur under total vision loss, out of distribution (OOD) textures, and periodic dropout Under total vision loss, the scan raises success from 45.2% to 64.6%, while baselines reach 44.6% to 56.6% with the scan on Under out of distribution textures, it is the only LiDAR-capable model whose success improves with the scan, rising from 44.2% to 58.4% Under periodic vision dropout at 25% and 50% duty, FLYNN-LiDAR achieves 93.6% and 89.6% success with fewer collisions Comparative Analysis and Efficiency FLYNN-LiDAR shows significant wins against three LiDAR-capable baselines across 24 paired comparisons Against EfficientNet, MobileNet, and SmallWorld, FLYNN-LiDAR has 17 significant wins after Bonferroni correction Under full vision with LiDAR, FLYNN-LiDAR leads the best of the three LiDAR-capable baselines by 7.8 points Profiling evaluated forward paths yields 3.93M multiply accumulate operations per step for FLYNN-LiDAR, which is less than EfficientNet and MobileNet This suggests that while it shows gains, it does not achieve orders of magnitude advantage in computational efficiency compared to CNNs State Geometry and Limitations The near-linear sensory superposition persists with the scan, as the cosine is 0.9955 for FLYNN-LiDAR with LiDAR Under total vision loss, full-dimensional Fisher-style meandifference scores are 1.07 for FLYNN-LiDAR and 2.90 for SmallWorld However, the lateral effect is model-dependent, as MobileNet reverses the ordering of left/right off performance Limitations include evaluations in simulation with one planar robot and one connectome instance, meaning training-seed variance and the origin of lateral asymmetry remain unresolved The LiDAR-off condition for FLYNN-LiDAR is an ablation of a policy trained with the scan, not a separately trained baseline The energy estimates omit memory and sensor costs, and state-geometry analyses are correlational rather than causal A fixed 32-ray projection can add range input to the fly whisker pathway with only a learned scalar input gain FLYNN-LiDAR shows its strongest robustness under total vision loss, OOD textures, and periodic flicker Multi-seed training, realistic scan noise and refresh, and hardware evaluation are necessary next steps The project repository is https://github.com/Ali-Mahzarnia/flynn-lidar (public release at publication) REFERENCES 2611.03897 2611.03897 2611.03897 4 4 4 5 5 5 Page 1 --- 1 Fly Whiskers Meet LiDAR: Robust Navigation under Sensor Failure Ali Mahzarnia and Homayoun Jamshidi Abstract—We extend FLYNN, a leaky integrator recurrent network wired on the Drosophila melanogaster connectome, with a 32-ray 2D LiDAR interface that requires no learned fusion layer beyond a scalar input gain.

Ines: A 270 degree frontal scan at 8 m range is injected as inverted proximity into the 37 BM Vib head-bristle neurons.

Marcus: Across 24 paired comparisons in a fourstate vision matrix with LiDAR on or off, FLYNN-LiDAR has 17 significant wins, 5 nonsignificant comparisons, and 2 significant losses against three LiDAR-capable baselines.

Ines: Both losses are to MobileNet with the left eye off.

Paper summary: Ines: So we're wrapping up this look at "Fly Whiskers Meet LiDAR: Robust Navigation under Sensor Failure". Basically, these researchers extended a network based on the fly brain called FLYNN by adding a way to take in range data from a LiDAR scanner <ref:2611.03897#pg1>.

Marcus: Right, so the core idea is that you inject this fixed thirty-two-ray scan into the neuron pathway, and they found that this setup helps navigation when vision completely fails <ref:2611.03897#pg1>.

Ines: That's the big claim, right? They show it works best when you lose all vision, or when textures are totally weird, or even if there’s a constant flicker in the input <ref:2611.03897#pg1>.

Marcus: And the numbers back that up—under total loss of vision, their success rate jumps from forty five point two percent to sixty four point six percent compared to just using the network alone <ref:2611.03897#pg1>.

Ines: It suggests that this range input is really useful for maintaining control when primary visual data disappears. What does this mean for how we think about these tiny robots?

Marcus: For us, it means we can build systems that don't rely on a single type of sensor being perfect; having multiple ways to sense the world makes the whole thing way more stable <ref:2611.03897#pg1>.

Marcus: So, while it’s a solid addition to the fly connectome architecture, we need to see if those gains hold up when we move this into real hardware with noisy data <ref:2611.03897#pg1>.

Conclusion: Ines: So we're wrapping up our look at "Fly Whiskers Meet LiDAR: Robust Navigation under Sensor Failure". The main point here is that they took an existing network, FLYNN, and successfully gave it a way to use range data from a LiDAR scanner without needing any extra fusion layer beyond just a learned scalar input gain.

Marcus: Exactly. They showed that injecting this fixed thirty-two-ray scan into the fly whisker pathway helps navigation when vision completely fails or when the textures are totally weird.

Ines: That's what they claim, right? They found success rates jumping from forty five point two percent to sixty four point six percent under total vision loss.

Marcus: That’s a big jump for those failure conditions because the baselines didn't improve much at all when vision went down.

Ines: It means that this specific range input isn't just extra data; it fundamentally changes how the network handles sensory input under stress.

Marcus: And while they show this is robust, they also point out that the gains aren't huge in terms of raw speed compared to those other models when you look at the operations per step.

Ines: So, while it’s a solid addition to the fly connectome architecture and definitely improves resilience, we still need to see if those gains hold up when we move this into real hardware with noisy data.

Marcus: That's what we need to figure out next—how does this actually perform when it’s running in the messy world.

q-bio.NC

Submitted: 2026-11-02

Updated: 2026-11-17

Project page: https://ali-mahzarnia.github.io/flynn-lidar

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 75/100

The gist: Fly Whiskers Meet LiDAR: Robust Navigation under Sensor Failure extends FLYNN by integrating a 32-ray 2D LiDAR interface into a fly connectome network to improve navigation when vision fails The…

Key concepts

FLYNN
A recurrent neural network designed for robot navigation based on the topology of the fly brain. It processes sensory input through a series of interconnected neurons that learn how to navigate environments by integrating information across these biological pathways.
LiDAR Interface
A mechanism to introduce 32-ray 2D LiDAR scan data into the fly's neural network. This is done by projecting the scan onto specific head-bristle neurons with fixed weights and a single learned scalar gain, allowing the network to use range information.
Total Vision Loss Robustness
The ability of FLYNN-LiDAR to maintain high navigation success even when visual input is completely lost. The LiDAR scan provides crucial spatial context during these failures, significantly increasing performance compared to models relying solely on vision.

Terminology

Summary

Fly Whiskers Meet LiDAR: Robust Navigation under Sensor Failure extends FLYNN by integrating a 32-ray 2D LiDAR interface into a fly connectome network to improve navigation when vision fails The strongest benefits occur under total vision loss, out of distribution (OOD) textures, and periodic dropout FLYNN-LiDAR achieves 96. p. 0

Where the paper says it

1 Fly Whiskers Meet LiDAR: Robust Navigation under Sensor Failure Ali Mahzarnia and Homayoun Jamshidi Abstract—We extend FLYNN, a leaky integrator recurrent network wired on the Drosophila melanogaster connectome, with a 32-ray 2D LiDAR interface that requires no learned fusion layer beyond a scalar input gain. A 270 degree frontal scan at 8 m range is injected as inverted proximity into the 37 BM Vib head-bristle neurons.

6% success under full vision, whereas the baselines reach 44. p. 0
Where the paper says it

Under total vision loss, the scan raises success from 45. 6%; the baselines reach 44. p. 0

Where the paper says it

Under total vision loss, the scan raises success from 45. 6%; the baselines reach 44.

6% to 56.6% with the scan on The result is not universal superiority, as MobileNet retains a significant left-eye-off advantage

Sensor Integration and Mechanism

The LiDAR interface involves injecting a fixed, spatially ordered projection of a 32-ray 2D LiDAR scan into the 37 BM Vib head-bristle neurons This projection has fixed weights and one learned scalar input gain, which is integrated in the recurrent network The scan is represented by encoding each continuous range as 1 − r/8, where zero denotes free space and one denotes contact Rays are assigned by hemifield to the nearest of 18 left and 19 right BM Vib afferents in a fixed midline-to-lateral map The LiDAR-to-whisker mapping preserves true anatomical laterality, unlike the small-world control where laterality is assigned by convention

Performance Under Stress Conditions

The robustness of FLYNN-LiDAR is demonstrated across several failure modes and conditions The strongest benefits occur under total vision loss, out of distribution (OOD) textures, and periodic dropout Under total vision loss, the scan raises success from 45. p. 0

Where the paper says it

6% success. p. 0

Where the paper says it

With full vision, FLYNNLiDAR reaches 96. 6% success.

Under total vision loss, the scan raises success from 45.

2% to 64.6%, while baselines reach 44. p. 0
Where the paper says it

Under total vision loss, the scan raises success from 45. 6%; the baselines reach 44.

6% to 56.6% with the scan on Under out of distribution textures, it is the only LiDAR-capable model whose success improves with the scan, rising from 44. p. 0
Where the paper says it

6% with the scan on. Under out of distribution textures, it is the only LiDAR-capable model whose success improves with the scan (44.

2% to 58.4% Under periodic vision dropout at 25% and 50% duty, FLYNN-LiDAR achieves 93. p. 0
Where the paper says it

Under periodic vision dropout, FLYNN-LiDAR achieves 93. 6% success at 25% and 50% duty with fewer collisions.

6% and 89.6% success with fewer collisions

Comparative Analysis and Efficiency

FLYNN-LiDAR shows significant wins against three LiDAR-capable baselines across 24 paired comparisons Against EfficientNet, MobileNet, and SmallWorld, FLYNN-LiDAR has 17 significant wins after Bonferroni correction Under full vision with LiDAR, FLYNN-LiDAR leads the best of the three LiDAR-capable baselines by 7.8 points Profiling evaluated forward paths yields 3. p. 5

Where the paper says it

FLYNN-LiDAR shows its strongest robustness under total vision loss, OOD textures, and periodic flicker, while MobileNet retains a left-eye-specific advantage and the lateral effect remains unexplained. Profiling the evaluated paths yields 4.

93M multiply accumulate operations per step for FLYNN-LiDAR, which is less than EfficientNet and MobileNet This suggests that while it shows gains, it does not achieve orders of magnitude advantage in computational efficiency compared to CNNs

State Geometry and Limitations

The near-linear sensory superposition persists with the scan, as the cosine is 0.9955 for FLYNN-LiDAR with LiDAR Under total vision loss, full-dimensional Fisher-style meandifference scores are 1. p. 3

Where the paper says it

Under total vision loss, full-dimensional Fisher-style meandifference scores are 1. 07 for FLYNN-LiDAR and 2. p. 0

Where the paper says it

1 Fly Whiskers Meet LiDAR: Robust Navigation under Sensor Failure Ali Mahzarnia and Homayoun Jamshidi Abstract—We extend FLYNN, a leaky integrator recurrent network wired on the Drosophila melanogaster connectome, with a 32-ray 2D LiDAR interface that requires no learned fusion layer beyond a scalar input gain. A 270 degree frontal scan at 8 m range is injected as inverted proximity into the 37 BM Vib head-bristle neurons.

07 for FLYNN-LiDAR and 2.90 for SmallWorld However, the lateral effect is model-dependent, as MobileNet reverses the ordering of left/right off performance Limitations include evaluations in simulation with one planar robot and one connectome instance, meaning training-seed variance and the origin of lateral asymmetry remain unresolved The LiDAR-off condition for FLYNN-LiDAR is an ablation of a policy trained with the scan, not a separately trained baseline The energy estimates omit memory and sensor costs, and state-geometry analyses are correlational rather than causal

Conclusion

A fixed 32-ray projection can add range input to the fly whisker pathway with only a learned scalar input gain FLYNN-LiDAR shows its strongest robustness under total vision loss, OOD textures, and periodic flicker Multi-seed training, realistic scan noise and refresh, and hardware evaluation are necessary next steps The project repository is https://github.com/Ali-Mahzarnia/flynn-lidar (public release at publication)

REFERENCES

2611.03897

2611.03897

2611.03897

4

4

4

5

5

5

--- Page 1 --- 1 Fly Whiskers Meet LiDAR: Robust Navigation under Sensor Failure Ali Mahzarnia and Homayoun Jamshidi Abstract—We extend FLYNN, a leaky integrator recurrent network wired on the Drosophila melanogaster connectome, with a 32-ray 2D LiDAR interface that requires no learned fusion layer beyond a scalar input gain. A 270 degree frontal scan at 8 m range is injected as inverted proximity into the 37 BM Vib head-bristle neurons. p. 0

Where the paper says it

1 Fly Whiskers Meet LiDAR: Robust Navigation under Sensor Failure Ali Mahzarnia and Homayoun Jamshidi Abstract—We extend FLYNN, a leaky integrator recurrent network wired on the Drosophila melanogaster connectome, with a 32-ray 2D LiDAR interface that requires no learned fusion layer beyond a scalar input gain. A 270 degree frontal scan at 8 m range is injected as inverted proximity into the 37 BM Vib head-bristle neurons.

Across 24 paired comparisons in a fourstate vision matrix with LiDAR on or off, FLYNN-LiDAR has 17 significant wins, 5 nonsignificant comparisons, and 2 significant losses against three LiDAR-capable baselines. p. 0
Where the paper says it

A 270 degree frontal scan at 8 m range is injected as inverted proximity into the 37 BM Vib head-bristle neurons. Across 24 paired comparisons in a fourstate vision matrix with LiDAR on or off, FLYNN-LiDAR has 17 significant wins, 5 nonsignificant comparisons, and 2 significant losses against three LiDAR-capable baselines.

Both losses are to MobileNet with the left eye off. p. 0
Where the paper says it

Across 24 paired comparisons in a fourstate vision matrix with LiDAR on or off, FLYNN-LiDAR has 17 significant wins, 5 nonsignificant comparisons, and 2 significant losses against three LiDAR-capable baselines. Both losses are to MobileNet with the left eye off.

With full vision, FLYNNLiDAR reaches 96. p. 0
Where the paper says it

Both losses are to MobileNet with the left eye off. With full vision, FLYNNLiDAR reaches 96.

6% success. Under total vision loss, the scan raises success from 45.2% to 64.6%; the baselines reach 44.6% to 56.6% with the scan on Profiling the evaluated forward paths gives 3. p. 0
Where the paper says it

6% success at 25% and 50% duty with fewer collisions. Profiling the evaluated forward paths gives 3.

93M multiply accumulate operations per step; EfficientNet and MobileNet require 4. p. 0
Where the paper says it

Profiling the evaluated forward paths gives 3. 93M multiply accumulate operations per step; EfficientNet and MobileNet require 4.

60× and 2.94× as many Finalstate analysis shows near-linear sensory superposition (cosine 0. p. 0
Where the paper says it

94× as many. Finalstate analysis shows near-linear sensory superposition (cosine 0.

9955) Index Terms—Robot navigation, connectome-based neural networks, sensor fault tolerance, LiDAR, neuromorphic computing

--- Page 2 --- 2 Fig. 1. Sensory suite: arena view with the 32-ray 270 degree LiDAR fan, left and right camera frames, and the inverted scan. p. 1

Where the paper says it

Sensory suite: arena view with the 32-ray 270 degree LiDAR fan, left and right camera frames, and the inverted scan. The fixed projection maps the rays onto 37 BM Vib afferents (18 left, 19 right).

The fixed projection maps the rays onto 37 BM Vib afferents (18 left, 19 right). p. 1
Where the paper says it

Sensory suite: arena view with the 32-ray 270 degree LiDAR fan, left and right camera frames, and the inverted scan. The fixed projection maps the rays onto 37 BM Vib afferents (18 left, 19 right).

In the connectome model, the LiDAR-to-whisker mapping preserves true anatomical laterality; in the small-world control, laterality is assigned by convention, as the graph carries no spatial anatomy. p. 1
Where the paper says it

BM Vib does not overlap the separate tactile head-bristle input. In the connectome model, the LiDAR-to-whisker mapping preserves true anatomical laterality; in the small-world control, laterality is assigned by convention, as the graph carries no spatial anatomy.

The control receives 37 arbitrarily labeled sites with the same 18/19 split. p. 1
Where the paper says it

In the connectome model, the LiDAR-to-whisker mapping preserves true anatomical laterality; in the small-world control, laterality is assigned by convention, as the graph carries no spatial anatomy. The control receives 37 arbitrarily labeled sites with the same 18/19 split.

Each CNN receives the identical 32-dimensional vector as additional input. p. 1
Where the paper says it

The control receives 37 arbitrarily labeled sites with the same 18/19 split. Each CNN receives the identical 32-dimensional vector as additional input.

Fig. 1 shows the sensory suite

--- Page 3 --- MAHZARNIA AND JAMSHIDI: FLY WHISKERS MEET LIDAR: ROBUST NAVIGATION UNDER SENSOR FAILURE A. Mahzarnia and H. p. 0

Where the paper says it

1 Fly Whiskers Meet LiDAR: Robust Navigation under Sensor Failure Ali Mahzarnia and Homayoun Jamshidi Abstract—We extend FLYNN, a leaky integrator recurrent network wired on the Drosophila melanogaster connectome, with a 32-ray 2D LiDAR interface that requires no learned fusion layer beyond a scalar input gain. A 270 degree frontal scan at 8 m range is injected as inverted proximity into the 37 BM Vib head-bristle neurons.

p. 0
Where the paper says it

1 Fly Whiskers Meet LiDAR: Robust Navigation under Sensor Failure Ali Mahzarnia and Homayoun Jamshidi Abstract—We extend FLYNN, a leaky integrator recurrent network wired on the Drosophila melanogaster connectome, with a 32-ray 2D LiDAR interface that requires no learned fusion layer beyond a scalar input gain. A 270 degree frontal scan at 8 m range is injected as inverted proximity into the 37 BM Vib head-bristle neurons.

Jamshidi are independent researchers (e-mail: mhznariman. p. 0
Where the paper says it

Mahzarnia and H. Jamshidi are independent researchers (e-mail: mhznariman.

p. 0
Where the paper says it

Mahzarnia and H. Jamshidi are independent researchers (e-mail: mhznariman.

ali@gmail.com; homayoun. p. 0
Where the paper says it

Jamshidi are independent researchers (e-mail: mhznariman. com; homayoun.

p. 0
Where the paper says it

Jamshidi are independent researchers (e-mail: mhznariman. com; homayoun.

jamshidi@gmail. p. 0
Where the paper says it

com; homayoun. jamshidi@gmail.

p. 0
Where the paper says it

com; homayoun. jamshidi@gmail.

com). Both authors contributed equally to this work A fixed 32-ray to 37-neuron projection through the fly BM Vib whisker pathway, with true anatomical laterality and only a learned scalar input gain

--- Page 4 --- D. Evaluation protocol Experiments use MuJoCo [12], a 14 by 14 m arena, 20 random obstacles, and a random goal. p. 1

Where the paper says it

Realized episode and optimizer budgets differ from the nominal schedule for every model and are listed in Section V-B. Evaluation protocol Experiments use MuJoCo [12], a 14 by 14 m arena, 20 random obstacles, and a random goal.

p. 1
Where the paper says it

Realized episode and optimizer budgets differ from the nominal schedule for every model and are listed in Section V-B. Evaluation protocol Experiments use MuJoCo [12], a 14 by 14 m arena, 20 random obstacles, and a random goal.

Success is final goal distance below 0. p. 1
Where the paper says it

Evaluation protocol Experiments use MuJoCo [12], a 14 by 14 m arena, 20 random obstacles, and a random goal. Success is final goal distance below 0.

p. 1
Where the paper says it

Evaluation protocol Experiments use MuJoCo [12], a 14 by 14 m arena, 20 random obstacles, and a random goal. Success is final goal distance below 0.

8 m. Episodes have at most 600 steps at dt = 0. p. 1
Where the paper says it

Success is final goal distance below 0. Episodes have at most 600 steps at dt = 0.

02 s The matrix crosses full vision, left eye off, right eye off, and total loss with LiDAR on or off Disabled cameras are zeroed

--- Page 5 --- E. Implementation details The recurrence is ht+1 = aht+(1−a) tanh(W ht+xt+b), with learned retention a per cell class [1]. p. 1

Where the paper says it

Nonsignificance is not treated as evidence of equivalence. Implementation details The recurrence is ht+1 = aht+(1−a) tanh(W ht+xt+b), with learned retention a per cell class [1].

In the evaluated FLYNN checkpoints, 5,342,446 connection records coalesce to 3,732,460 nonzero recurrent weights. p. 1
Where the paper says it

Implementation details The recurrence is ht+1 = aht+(1−a) tanh(W ht+xt+b), with learned retention a per cell class [1]. In the evaluated FLYNN checkpoints, 5,342,446 connection records coalesce to 3,732,460 nonzero recurrent weights.

Including these weights, 138,584 biases gives 4. p. 1
Where the paper says it

In the evaluated FLYNN checkpoints, 5,342,446 connection records coalesce to 3,732,460 nonzero recurrent weights. Including these weights, 138,584 biases, 8,542 cell-class leak logits, readout and wind MLPs, and scalar gains gives 4.

08M learned parameters for FLYNN-LiDAR and FLYNN (repro. p. 1
Where the paper says it

Including these weights, 138,584 biases, 8,542 cell-class leak logits, readout and wind MLPs, and scalar gains gives 4. 08M learned parameters for FLYNN-LiDAR and FLYNN (repro.

) SmallWorld has 138,603 nodes MobileNet and EfficientNet have 3. p. 1
Where the paper says it

08M learned parameters for FLYNN-LiDAR and FLYNN (repro. SmallWorld has 138,603 nodes, 5,405,517 recurrent weights, 1,303 rather than 1,288 readout inputs, and 5.

99M and 5.29M parameters, respectively

--- Page 6 --- VI. CONCLUSION A fixed 32-ray projection can add range input to the fly whisker pathway with only a learned scalar input gain FLYNN-LiDAR shows its strongest robustness under total vision loss, OOD textures, and periodic flicker while MobileNet retains a left-eye-specific advantage and the lateral effect remains unexplained Profiling the evaluated paths yields 4. p. 5

Where the paper says it

CONCLUSION A fixed 32-ray projection can add range input to the fly whisker pathway with only a learned scalar input gain. FLYNN-LiDAR shows its strongest robustness under total vision loss, OOD textures, and periodic flicker, while MobileNet retains a left-eye-specific advantage and the lateral effect remains unexplained.

60× and 2.94× per-step MAC ratios against the CNNs, with lower estimated episode energy but not an orders-of-magnitude advantage Memory access, sensor acquisition, and sensor power are excluded; sparse and dense memory traffic could change the ratios in either direction, so the MAC ratios are not bounds on total energy

--- Page 7 --- REFERENCES [1] B. Wang and J. p. 5

Where the paper says it

REFERENCES [1] B. Wang and J.

p. 5
Where the paper says it

REFERENCES [1] B. Wang and J.

Chen, “FLYNN: Robust neural network for robot navigation using fly brain topology,” 2026, arXiv:2607. p. 5
Where the paper says it

Wang and J. Chen, “FLYNN: Robust neural network for robot navigation using fly brain topology,” 2026, arXiv:2607.

p. 5
Where the paper says it

Wang and J. Chen, “FLYNN: Robust neural network for robot navigation using fly brain topology,” 2026, arXiv:2607.

00025. [2] S. Ross, G. Gordon, and D. p. 5
Where the paper says it

Chen, “FLYNN: Robust neural network for robot navigation using fly brain topology,” 2026, arXiv:2607. Gordon, and D.

p. 5
Where the paper says it

Chen, “FLYNN: Robust neural network for robot navigation using fly brain topology,” 2026, arXiv:2607. Gordon, and D.

Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” in Proc. p. 5
Where the paper says it

Gordon, and D. Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” in Proc.

p. 5
Where the paper says it

Gordon, and D. Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” in Proc.

14th Int. Conf. Artif. Intell., 2011, pp. p. 5
Where the paper says it

Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” in Proc. , 2011, pp.

p. 5
Where the paper says it

Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” in Proc. , 2011, pp.

627–635

--- Page 8 --- [3] J. K. Lappalainen, F. p. 5

Where the paper says it

, 2011, pp. Lappalainen, F.

D. Tschopp, S. p. 5
Where the paper says it

Lappalainen, F. Tschopp, S.

-C Yu et al., “Connectome-constrained networks predict neural activity across the fly visual system,” Nature, vol. p. 5
Where the paper says it

Macke, and S. Turaga, “Connectome-constrained networks predict neural activity across the fly visual system,” Nature, vol.

634, no. 8032, pp. 1132–1140, 2024 [4] P. p. 5
Where the paper says it

Turaga, “Connectome-constrained networks predict neural activity across the fly visual system,” Nature, vol. 1132–1140, 2024.

K. Shiu et al., “A Drosophila computational brain model reveals sensorimotor processing,” Nature, vol. p. 5
Where the paper says it

Sandoval, J. , “A Drosophila computational brain model reveals sensorimotor processing,” Nature, vol.

634, no. 8032, pp. 210–219, 2024 [5] N. p. 5
Where the paper says it

, “A Drosophila computational brain model reveals sensorimotor processing,” Nature, vol. 210–219, 2024.

Patel et al., “A deep learning gated architecture for UGV navigation robust to sensor failures,” Robotics and Autonomous Systems, vol. p. 5
Where the paper says it

Krishnamurthy, and F. Khorrami, “A deep learning gated architecture for UGV navigation robust to sensor failures,” Robotics and Autonomous Systems, vol.

116, pp. 80–97, 2019 [6] M. p. 5
Where the paper says it

Khorrami, “A deep learning gated architecture for UGV navigation robust to sensor failures,” Robotics and Autonomous Systems, vol. 80–97, 2019.

Chahine et al. p. 5
Where the paper says it

80–97, 2019. Chahine, R.

, “Robust flight navigation out of distribution with liquid neural networks,” Science Robotics, vol. p. 5
Where the paper says it

Amini, and D. Rus, “Robust flight navigation out of distribution with liquid neural networks,” Science Robotics, vol.

8, no. 77, p. eadc8892

--- Page 9 --- [7] D. J. Watts and S. p. 5

Where the paper says it

eadc8892, 2023. Watts and S.

p. 5
Where the paper says it

eadc8892, 2023. Watts and S.

H. Strogatz, “Collective dynamics of ‘small-world’ networks,” Nature, vol. p. 5
Where the paper says it

Watts and S. Strogatz, “Collective dynamics of ‘small-world’ networks,” Nature, vol.

p. 5
Where the paper says it

Watts and S. Strogatz, “Collective dynamics of ‘small-world’ networks,” Nature, vol.

393, no. 6684, pp. 440–442, 1998 [8] S. p. 5
Where the paper says it

Strogatz, “Collective dynamics of ‘small-world’ networks,” Nature, vol. 440–442, 1998.

Dorkenwald et al. p. 5
Where the paper says it

440–442, 1998. Dorkenwald, A.

p. 5
Where the paper says it

440–442, 1998. Dorkenwald, A.

, “Neuronal wiring diagram of an adult brain: Supplementary data,” Nature, vol. p. 5
Where the paper says it

Dorkenwald et al. , “Neuronal wiring diagram of an adult brain: Supplementary data,” 2024.

p. 5
Where the paper says it

Dorkenwald et al. , “Neuronal wiring diagram of an adult brain: Supplementary data,” 2024.

634, no. 8032, pp. 124–138

--- Page 10 --- [9] P. Schlegel et al. p. 5

Where the paper says it

Sterling, P. Schlegel, S.

p. 5
Where the paper says it

Sterling, P. Schlegel, S.

, “Whole-brain annotation and multi-connectome cell typing of Drosophila,” Nature, vol. p. 5
Where the paper says it

Munnelly et al. , “Whole-brain annotation and multi-connectome cell typing of Drosophila,” Nature, vol.

p. 5
Where the paper says it

Munnelly et al. , “Whole-brain annotation and multi-connectome cell typing of Drosophila,” Nature, vol.

634, no. 8032, pp. 139–152

--- Page 11 --- [10] SLAMTEC, “RPLIDAR A1 laser range scanner specifications,” https://www.slamtec.com/en/lidar/a1spec, 2026, accessed: October 24, 2026

--- Page 13 --- [11] I. Ulrich and J. p. 5

Where the paper says it

com/en/lidar/a1spec, 2026, accessed: October 24, 2026. Ulrich and J.

Borenstein, “VFH: Local obstacle avoidance with lookahead verification,” in Proc. p. 5
Where the paper says it

Ulrich and J. Borenstein, “VFH*: Local obstacle avoidance with lookahead verification,” in Proc.

2000 ICRA. IEEE Int. Conf. Robot. Autom., vol. 3, 2000, pp. p. 5
Where the paper says it

Borenstein, “VFH*: Local obstacle avoidance with lookahead verification,” in Proc. 3, 2000, pp.

2505–2511

--- Page 14 --- [12] E. Todorov and T. p. 5

Where the paper says it

2505–2511. Todorov, T.

p. 5
Where the paper says it

2505–2511. Todorov, T.

Erez and Y. p. 5
Where the paper says it

Todorov, T. Erez, and Y.

p. 5
Where the paper says it

Todorov, T. Erez, and Y.

Tassa, “MuJoCo: A physics engine for model-based control,” in 2012 IEEE/RSJ Int. p. 5
Where the paper says it

Erez, and Y. Tassa, “MuJoCo: A physics engine for model-based control,” in 2012 IEEE/RSJ Int.

p. 5
Where the paper says it

Erez, and Y. Tassa, “MuJoCo: A physics engine for model-based control,” in 2012 IEEE/RSJ Int.

Conf. Intell. Robots Syst. p. 5
Where the paper says it

Tassa, “MuJoCo: A physics engine for model-based control,” in 2012 IEEE/RSJ Int. Robots Syst.

p. 5
Where the paper says it

Tassa, “MuJoCo: A physics engine for model-based control,” in 2012 IEEE/RSJ Int. Robots Syst.

, 2012, pp. p. 5
Where the paper says it

Erez, and Y. Tassa, “MuJoCo: A physics engine for model-based control,” in 2012 IEEE/RSJ Int.

p. 5
Where the paper says it

Erez, and Y. Tassa, “MuJoCo: A physics engine for model-based control,” in 2012 IEEE/RSJ Int.

5026–5033

--- Page 15 --- [13] A. Howard et al., “Searching for MobileNetV3,” in Proc. p. 5

Where the paper says it

Vasudevan, Q. Adam, “Searching for MobileNetV3,” in Proc.

IEEE/CVF Int. p. 5
Where the paper says it

Adam, “Searching for MobileNetV3,” in Proc. IEEE/CVF Int.

Conf. Comput. Vis., 2019, pp. p. 5
Where the paper says it

Khorrami, “A deep learning gated architecture for UGV navigation robust to sensor failures,” Robotics and Autonomous Systems, vol. 80–97, 2019.

p. 5
Where the paper says it

Khorrami, “A deep learning gated architecture for UGV navigation robust to sensor failures,” Robotics and Autonomous Systems, vol. 80–97, 2019.

6105–6114 [14] M. p. 5
Where the paper says it

, 2019, pp. 6105–6114.

Tan and Q. Le, “EfficientNet: Rethinking model scaling for convolutional neural networks,” in Proc. p. 5
Where the paper says it

(ICCV), 2019. Le, “EfficientNet: Rethinking model scaling for convolutional neural networks,” in Proc.

36th Int. Conf. Mach Learn., 2019, pp. 6105–6114

--- Page 20 --- [15] M. Horowitz, “1. p. 5

Where the paper says it

6105–6114. Horowitz, “1.

1 computing’s energy problem (and what we can do about it),” in 2014 IEEE Int. p. 5
Where the paper says it

Horowitz, “1. 1 computing’s energy problem (and what we can do about it),” in 2014 IEEE Int.

Solid-State Circuits Conf. p. 5
Where the paper says it

1 computing’s energy problem (and what we can do about it),” in 2014 IEEE Int. Solid-State Circuits Conf.

Dig. Tech. Papers (ISSCC), 2014, pp. p. 5
Where the paper says it

Solid-State Circuits Conf. Papers (ISSCC), 2014, pp.

10–14

--- Page 26 --- [7] D. J. Watts and S. H. Strogatz, “Collective dynamics of ‘small-world’ networks,” Nature, vol. 393, no. 6684, pp. 440–442, 1998

--- Page 27 --- [8] S. Dorkenwald et al., “Neuronal wiring diagram of an adult brain: Supplementary data,” Nature, vol. 634, no. 8032, pp. 124–138

--- Page 27 --- [9] P. Schlegel et al., “Whole-brain annotation and multi-connectome cell typing of Drosophila,” Nature, vol. 634, no. 8032, pp. 139–152

--- Page 27 --- [10] SLAMTEC, “RPLIDAR A1 laser range scanner specifications,” https://www.slamtec.com/en/lidar/a1spec, 2026, accessed: October 24, 2026

--- Page 3 --- MAHZARNIA AND JAMSHIDI: FLY WHISKERS MEET LIDAR: ROBUST NAVIGATION UNDER SENSOR FAILURE A. Mahzarnia and H. Jamshidi are independent researchers (e-mail: mhznariman.ali@gmail.com; homayoun.jamshidi@gmail.com). Both authors contributed equally to this work

--- Page 4 --- D. Evaluation protocol Experiments use MuJoCo [12], a 14 by 14 m arena, 20 random obstacles, and a random goal. Success is final goal distance below 0.8 m Episodes have at most 600 steps at dt = 0. p. 1

Where the paper says it

Success is final goal distance below 0. Episodes have at most 600 steps at dt = 0.

02 s The matrix crosses full vision, left eye off, right eye off, and total loss with LiDAR on or off Disabled cameras are zeroed

--- Page 7 --- REFERENCES [1] B. Wang and J. Chen, “FLYNN: Robust neural network for robot navigation using fly brain topology,” 2026, arXiv:2607.00025. [2] S. Ross, G. Gordon, and D. Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” in Proc. 14th Int. Conf. Artif., 2011, pp. 627–635

--- Page 14 --- [12] E. Todorov and T. Erez and Y. Tassa, “MuJoCo: A physics engine for model-based control,” in 2012 IEEE/RSJ Int. Conf. Intell. Robots Syst., 2012, pp.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Enhanced sensor fault tolerance in connectome-based models by injecting a fixed 32-ray 2D LiDAR scan into head bristle neurons, which requires no learned fusion layer beyond a scalar input gain. This allows the network to improve navigation when vision fails, achieving success rates up to 96.6% under full vision.

  2. Improved performance under challenging conditions by demonstrating that The strongest benefits occur under total vision loss, out of distribution (OOD) textures, and periodic dropout, showing FLYNN-LiDAR's success improves from 45.2% to 64.6% compared to baselines that drop significantly.

  3. Better handling of texture shifts by using LiDAR input, as Under out of distribution textures, it is the only LiDAR-capable model whose success improves with the scan (44.2% to 58.4%), whereas CNNs score 0%.

  4. Increased collision avoidance capability by showing that under total vision loss, the scan raises success from 45.2% to 64.6%; the baselines reach 44.6% to 56.6% with the scan on, and that FLYNN-LiDAR has 3.3× and 3.6× fewer collisions than the CNNs at higher success.

  5. More robust response to vision flicker by achieving 93.6% and 89.6% success at 25% and 50% duty with periodic vision dropout, outperforming baselines significantly under these conditions.

Abstract

We extend FLYNN, a leaky integrator recurrent network wired on the Drosophila melanogaster connectome, with a 32-ray 2D LiDAR interface that requires no learned fusion layer beyond a scalar input gain. A 270 degree frontal scan at 8 m range is injected as inverted proximity into the 37 BM Vib head-bristle neurons. Across 24 paired comparisons in a four- state vision matrix with LiDAR on or off, FLYNN-LiDAR has 17 significant wins, 5 nonsignificant comparisons, and 2 significant losses against three LiDAR-capable baselines. Both losses are to MobileNet with the left eye off. With full vision, FLYNN- LiDAR reaches 96.6% success. Under total vision loss, the scan raises success from 45.2% to 64.6%; the baselines reach 44.6% to 56.6% with the scan on. Under out of distribution textures, it is the only LiDAR-capable model whose success improves with the scan (44.2% to 58.4%); both CNNs score 0%, and SmallWorld declines from 26.8% to 7.4%. Under periodic vision dropout, FLYNN-LiDAR achieves 93.6% and 89.6% success at 25% and 50% duty with fewer collisions. Profiling the evaluated forward paths gives 3.93M multiply accumulate operations per step; EfficientNet and MobileNet require 4.60x and 2.94x as many. Final- state analysis shows near-linear sensory superposition (cosine 0.9955).

Sources

Related papers