Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 158 of the 159 references it cites were found in a registry or an index, and 1 are not checked yet; 36 of the 82 sources its claims cite were checked for support.

NeuroMechFly v2: simulating embodied sensorimotor control in adult Drosophila

Open access

Published version of a preprint

Abstract

We present NeuroMechFly v2, a framework that expands Drosophila neuromechanical modeling by enabling visual and olfactory sensing, ascending motor feedback, and complex terrains that can be navigated using leg adhesion. We illustrate its capabilities by first constructing biologically inspired locomotor controllers that use ascending motor feedback to perform path integration and head stabilization. Then, we add visual and olfactory sensing to this controller and train it using reinforcement learning to perform a multimodal navigation task in closed loop. Finally, we illustrate more biorealistic modeling in two ways: our model navigates a complex odor plume using a Drosophila odor taxis strategy, and it uses a connectome-constrained visual system network to follow another simulated fly. With this framework, NeuroMechFly can be used to accelerate the discovery of explanatory models of the nervous system and to develop machine learning-based controllers for autonomous artificial agents and robots.

Verified 2026-10-12 · 1 reference not checked yet

Overview

CoChat Summary

TL;DR

NeuroMechFly 2.0 (FlyGym) expands the Drosophila neuromechanical model with MuJoCo physics, leg adhesion, visual/olfactory sensing, rugged terrain, and a Gymnasium-compatible interface for hierarchical sensorimotor control. It demonstrates terrain-dependent locomotor controller trade-offs and closed-loop multimodal navigation using a reinforcement-learning-trained controller.

Key findings

  • Leg adhesion (20–40 mN) enabled 3D locomotion: without adhesion the fly failed to sustain forward locomotion at slopes as small as 5°, while with adhesion it could locomote on inclines up to 120° and occasionally beyond 90°.
  • Benchmarking three controllers over flat, gapped, blocks, and mixed terrain (N = 20 trials each) showed the CPG controller was fast, the rule-based controller was more robust on rugged terrain, and the hybrid controller had higher baseline speed than the rule-based controller and covered greater distances on rugged terrain.
  • Vision and olfaction were implemented with 721 ommatidia per eye, a ~270° combined azimuth, ~17° binocular overlap, and a 7:3 yellow:pale ommatidia ratio; virtual odor sensors on antennae and maxillary palps supported visual object chasing and olfactory chemotaxis.
  • A hierarchical RL controller solved a multimodal task—avoiding a visual obstacle while approaching an attractive odor source over rugged terrain—using a vision model trained on 10,000 samples for 150 epochs and an RL decision module trained for 266,000 decision steps (~3.7 simulated hours).

Methods

  • FlyGym is a Python package built on MuJoCo and dm_control, complying with the Gymnasium MDP interface; actions include joint actuation signals and leg adhesion on/off, while observations include joint states, fly state, ground reaction forces, and optional visual/olfactory inputs.
  • The biomechanical model was updated with antennal segmentation, neck/head/leg adjustments, and leg adhesion via MuJoCo actuators; stepping kinematics were derived from recordings of tethered flies, and terrains included flat, gapped, blocks, and mixed surfaces.
  • Controllers included CPG-based, rule-based (Walknet-inspired), and hybrid CPG-plus-sensory-feedback strategies; turning was controlled by a two-dimensional descending signal, and the multimodal RL model used a graph-convolution vision module plus an SAC-trained MLP decision module.

Limitations

  • Stated: the model cannot fully represent real biological adhesion; lift-off mechanisms for Drosophila are unknown, so adhesion was abstracted as on during stance and off during swing; dorsal rim area ommatidia were not implemented; odor chemotaxis was simplified by the absence of complex plumes; and behaviors involving wings/halteres (flight), abdomen (egg-laying), and proboscis (feeding) were not yet demonstrated.
  • Stated future improvements: GPU-accelerated/differentiable simulation, connectome-constrained neural circuits, and musculoskeletal/tendon-muscle models.
  • Evident: visual color uses green and blue channels as proxies for yellow- and pale-type ommatidia rather than full spectral reflectance, and the RL navigation task used a simplified mixed terrain with reduced gap widths (0.2 mm) and block heights (0.3 mm); all reported results are simulation-only, with no real-fly validation of the full hierarchical controller presented.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41592-024-02497-y

Cites

158 of 159 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

8 Valid · 20 Not valid · 21 Not checked · 79 Own results and interpretations

50 more claims

Create a free account to see all 128 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 34
  • Probably not a substantial claim. 50
  • The quote only partly states this claim. 2

Create a free account to see these claims.

Citations

Cites

Show all 79 references

Research works printed but not linked: 3

  1. Ho, J. & Ermon, S. Generative adversarial imitation learning. In Proc. 30th International Conference on Neural Information Processing Systems. 4572–4580 (Curran Associates, 2016).
  2. Caggiano, V., Wang, H., Durandau, G., Sartori, M. & Kumar, V. MyoSuite: a contact-rich simulation suite for musculoskeletal motor control. Proc. Mach. Learn. Res. 168, 492–507 (2022).
  3. Holl, P., Thuerey, N. & Koltun, V. Learning to control PDEs with differentiable physics. In International Conference on Learning Representations (2020); https://openreview.net/forum?id=HyeSin4FPB

Cited by

See all 43

Metadata

Authors
Wang-Chen, S., Stimpfling, V. A., Lam, T. K. C., Özdil, P. G., Genoud, L., Hurtak, F., & Ramdya, P.
Affiliations
École Polytechnique Fédérale de Lausanne (CH)
Corresponding author
Ramdya, Pavan
Venue
Nature Methods
Published
2024-11-12
Volume
21
Issue
12
Pages
2353-2362
ISSN
  • 1548-7091 (ISSN-L)
  • 1548-7105
OpenAlex
W4404252550
Cited by
43
References
79
Citations by year
2026: 22 · 2025: 11 · 2024: 9
FWCI
7.32
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Methods (publishedVersion)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • Infoscience (Ecole Polytechnique Fédérale de Lausanne) (submittedVersion, cc-by-nc-nd)
Funders
  • Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung (175667, 181239)
  • Boehringer Ingelheim Fonds
  • Google
Topics
  • Life Sciences › Neuroscience › Cellular and Molecular Neuroscience › Neurobiology and Insect Physiology Research
  • Physical Sciences › Engineering › Control and Systems Engineering › Robotic Locomotion and Control
  • Physical Sciences › Engineering › Aerospace Engineering › Biomimetic flight and propulsion mechanisms
Keywords
Drosophila, sensorimotor control, path integration, head stabilization, multimodal navigation, reinforcement learning, robot locomotion
MeSH terms
Machine Learning, Animals, Behavior, Animal, Brain, Computer Simulation, Drosophila, Drosophila melanogaster, Models, Neurological, Smell, Connectome

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-12, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-12 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: NeuroMechFly v2: simulating embodied sensorimotor control in adult Drosophila. https://app.cochat.ai/papers/10.1038/s41592-024-02497-y
The paper
Wang-Chen, S., Stimpfling, V. A., Lam, T. K. C., Özdil, P. G., Genoud, L., Hurtak, F., & Ramdya, P. (2024). NeuroMechFly v2: simulating embodied sensorimotor control in adult Drosophila. Nature Methods. https://doi.org/10.1038/s41592-024-02497-y

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.