An Efficient Insect-inspired Approach for Visual Point-goal Navigation

TL;DR

This work addresses visual point-goal navigation by introducing an insect-inspired agent that fuses Central Complex (CX) path integration with Mushroom Body (MB) visual learning. The model uses online learning with multi-timescale memory consolidation to adaptively avoid obstacles and improve routes without any pretraining, achieving competitive performance with far lower computational cost than state-of-the-art RL/SLAM systems. Evaluations in Habitat and iGibson show robust obstacle avoidance, online path optimization, and resilience to motor perturbations, highlighting a practical, mapless navigation approach suitable for real robots. The study also discusses limitations and avenues for real-world deployment, including sim-to-real transfer, depth sensing, continual learning, and potential expansion to multi-agent systems.

Abstract

In this work we develop a novel insect-inspired agent for visual point-goal navigation. This combines abstracted models of two insect brain structures that have been implicated, respectively, in associative learning and path integration. We draw an analogy between the formal benchmark of the Habitat point-goal navigation task and the ability of insects to learn and refine visually guided paths around obstacles between a discovered food location and their nest. We demonstrate that the simple insect-inspired agent exhibits performance comparable to recent SOTA models at many orders of magnitude less computational cost. Testing in a more realistic simulated environment shows the approach is robust to perturbations.

Figures (12)

• Figure 1: Overview of our insect-inspired model. (A) The closed-loop learning and control of our model embodied in a virtual robot. The model is primarily composed of two insect brain-like modules, a mushroom body (MB) and a central complex (CX), They use three types of (preprocessed) data as inputs, odometry (I), collision (II), and vision (III) to calculate a desired target direction, which determines control commands to be sent to the robot (V). (B) The MB architecture. The architecture is primarily feedforward, consisting of projection neurons (PNs), Kenyon cells (KCs), and left and right MB output neurons (MBONs); there is also lateral inhibition amongst the KCs. The activation of the PNs encode preprocessed visual information (I), and the activation of the MBONs is the modulatory signal projected from the MB to the CX (IV). The PN-KC synapses are predetermined with random initialisation, and the KC-MBON synapses are plastic. In particular, their plasticity is gated by dopaminergic neurons (DANs), receiving reinforcement signals determined by the urge to escape (II). An escape urge is produced after collision detection with a punishment signal that fades away after a fixed duration with a reward signal following. Recent changes of KC-MBON weights are consolidated into long-term memory, given improved task performance; otherwise, the changes are discarded. (C) The angular calculation performed by the CX. A target direction is set, either towards the goal or to escape an obstacle, and then modulated by the MB signals (IV) to the left or right side, based on which MBON is more activated. The modulated target is then compared to the current orientation (I) to calculate the desired rotation.
• Figure 2: Model performance (SR (square) and SPL (cross), defined in §\ref{['sec-methods:metric']}) vs number of training frames. The results of our insect-inspired model (Full-first) and the ablated models (Odometry-only and Odometry-collision) were obtained from 100 independent test episodes in Habitat with the Gibson 4+ scenes by averaging the performance on the first trial for all test episodes. The tests were independent from each other with the full model's memory reset each time. The results for the Models 1-7 are taken from the reported results summarised in Table \ref{['tab:model-types']}. For fairer comparison, the performance of Model 1 is inflated deliberately by about a factor of $1.3$ from that in Table \ref{['tab:model-types']}, because mishkin2019benchmarking used a different simulator with a different dataset. The inflation factor is chosen based on the results in Table 2 of savva2019habitat, where different results of the same model were reported for both Gibson and Matterport3D scenes.
• Figure 3: Performance of insect-inspired models in Habitat and iGibson. The odometry-only model has both vision and collision pathways ablated, and the odometry-collision model only the vision pathway ablated. Full-first and Full-learnt represent the (same) full model's performance at the first trials and when it completed learning. The episodes could not be solved by the odometry-collision model were considered 'hard', and thus the model scored $0$.
• Figure 4: Summary of model collisions. Note the scales are different for the ablated vs full models.
• Figure 5: Performance of insect-inspired models with different memory consolidation mechanisms. Full-learnt: our insect-inspired model with selective memory consolidation. Full-excessive: the same model but with excessive, unconditional memory consolidation across all 20 trials. Full-hypothetical: a model memorising all trial experiences within an episode and cherrypicking the best performance.