Insect-inspired Approach for Point-goal Navigation

Overview

We develop a novel model for point-goal navigation by drawing an analogy between the formal benchmark of the Habitat point-goal navigation task and the ability of insects to discover, learn, and refine visually guided paths around obstacles between a discovered food location and their nest. The model consists of two components, abstracted from two insect brain structures, the Mushroom Body (MB) responsible for (visual) associative learning and the Central compleX (CX) for path integration.

To be more specific, the point-goal navigation task concerns if and how an agent can navigate from the start position to the goal position, given the coordinates of two positions, in an unknown environment with potential obstacles. We investigate, specifically, how the robot learns to avoid potential obstacles and optimise its route, using visual memory (with zero prior knowledge).

The MB-CX integrative model is tested in two simulators: Habitat and iGibson. Habitat is ued for comparing to state-of-the-art models, while the more realistic physics engine of iGibson permits validation of the model's robustness against perturbation.

Installation and Dependencies

Due to the different physics engines, the model needs to be run in the two different simulators. Two independet conda environments are recommended, because the two simulators require many common dependencies but of different versions, e.g. OpenCV. We have been using two independent environments throughout development and testing on an Ubuntu 22.04 LTS laptop with Python 3.8.13.

Initial steps to install the two simulators are shown below, and more details can be found in the corresponding README files in the Habitat and iGibson directories, respectively.

Habitat

1. Install habitat-sim (v0.3.1):
   1. Following README instructions for the conda option, but ignoring everything including and after Testing.
   2. Uninstall and (re)install numpy (v1.23) using pip, due to its compatiblitiy with habitat-lab.
2. Install habita-lab (v0.3.1):
   1. Following README instructions for the conda option, but not repeating the first steps as habitat-sim has been installed.
   2. Double-checking the versions of habitat-sim and habitat-lab, which should be the same.
3. Download assets under habitat-lab/data by git lfs pull.
   1. The download commands executed when installing habitat-sim and habitat-lab did not download the data; they only create pointers.
   2. At this stage, the jupyter notebooks under habitat-lab/examples should work.
   3. However, to make python files run, there is a conflict between opencv-python and PyQt5. The solution is to uninstall opencv-python and install opencv-python-headless instead, using pip.

iGibson

Please install all dependencies required by iGibson 2.0 first. Additionally with numpy and matplotlib, one can reproduce simulations and visualisations (given simulated data).

Reconstruction Datasets

After installing both simulators, one should be able to test run some scripts but in very few scenes. To fully replicate our work, at least the Habitat simulations in our paper, it is necessary to download the Gibson 4+ dataset, which requires an official license (free for research use).

The Insect-inspired Model

CX

The insect CX is a shallow neural circuit, playing a critical role in path integration, i.e. dead reckoning. To better align with the Habitat challenge, we abstract the CX into a simple algorithm for GPS+Compass.

MB

The insect MB, also a shallow circuit, is capable of rapid associative learning. The MB, as a visual memory in our models, is assumed to be a two-layered neural network, consisting of

• visual projection neurons (PN), receiving preprocessed visual inputs,
• Kenyon cells (KC), encoding any PN activity as a latent, sparse pattern, by multiplying the PN-KC weight matrix,
• MB output neurons (MBON), computing visual familiarity/novelty of the KC pattern, by multiplying the KC-MBON weight matrix.
• DopAminergic Neurons (DAN), activated by collision-related events and triggering learning. where
• the PN-KC matrix is randomly initialised to be binary and sparse, and fixed throughout a simulated experiment; and
• the KC-MBON matrix is initialised to be one, as learning is achieved by synaptic depression (i.e., decreasing weights, which is more consistent with the real MB).

Other than the specific learning rule, the most important parameters of the MB are N_pn (the number of PN), N_pn_perkc (the sparsity of the PN-KC connectivity), N_kc (the number of KC) and S_kc (the sparsity of KC activity).

Licenses

The MB-CX integrative model is different from the MB-only model in our previous work for visual route following, but we reuse part of the code from insect-inspired-route-following (under GNU GPL 3.0 license), particularly those for low-level perception and control in the iGibson simulators.

You are welcome to (re-)use everything in this repository, but please (re-)distribute any derivative works with the same GNU GPL 3.0 license and cite us:

1. Lu, Y., Cen, J., Alkhoury Maroun, R. et al. (2025). Insect-inspired Embodied Visual Route Following. Journal of Bionic Engineering. 22, 1167–1193 https://doi.org/10.1007/s42235-025-00695-8
2. Yihe, L., & Webb, B. (2026). An Efficient Insect-inspired Approach for Visual Point-goal Navigation. arXiv preprint arXiv:2601.16806.