A Bio-Inspired Research Paradigm of Collision Perception Neurons Enabling Neuro-Robotic Integration: The LGMD Case

TL;DR

This paper addresses efficient collision perception by leveraging locust LGMD neurons as a model for biologically plausible, low-power collision detection. It outlines two complementary computational approaches—the $\eta$-function single-neuron dynamics and multi-layer networks with image cues and inhibition—and discusses their real-world applicability and limitations. It surveys LGMD-based robotic implementations on ground and aerial platforms, including LGMD1, LGMD2, and hybrid models, and highlights sensor innovations such as event-based cameras. It argues for a bidirectional bio-inspired paradigm where neuroscience informs robotics and robotics provides testable neuroscience predictions, with future connections to connectomics and cross-species circuits.

Abstract

Compared to human vision, locust visual systems excel at rapid and precise collision detection, despite relying on only hundreds of thousands of neurons organized through a few neuropils. This efficiency makes them an attractive model system for developing artificial collision-detecting systems. Specifically, researchers have identified collision-selective neurons in the locust's optic lobe, called lobula giant movement detectors (LGMDs), which respond specifically to approaching objects. Research upon LGMD neurons began in the early 1970s. Initially, due to their large size, these neurons were identified as motion detectors, but their role as looming detectors was recognized over time. Since then, progress in neuroscience, computational modeling of LGMD's visual neural circuits, and LGMD-based robotics have advanced in tandem, each field supporting and driving the others. Today, with a deeper understanding of LGMD neurons, LGMD-based models have significantly improved collision-free navigation in mobile robots including ground and aerial robots. This review highlights recent developments in LGMD research from the perspectives of neuroscience, computational modeling, and robotics. It emphasizes a biologically plausible research paradigm, where insights from neuroscience inform real-world applications, which would in turn validate and advance neuroscience. With strong support from extensive research and growing application demand, this paradigm has reached a mature stage and demonstrates versatility across different areas of neuroscience research, thereby enhancing our understanding of the interconnections between neuroscience, computational modeling, and robotics. Furthermore, this paradigm would shed light upon the modeling and robotic research into other motion-sensitive neurons or neural circuits.

Figures (6)

• Figure 1: The reviewed bio-inspired research paradigm: neuroscience studies of the LGMD neuron, ranging from behavioral neuroscience and neuro-morphology, inform the development of computational models, such as single neuron modeling and multi-layered neural network, simulating LGMD's functionality and selectivity. These models are successfully implemented in robotics as embedded vision, such as micro-mobile robots and small quadcopters, for real-time collision detection and avoidance in navigation. The output and behavior of these LGMD-based embodiments, in turn, validate the models and provide meaningful feedback that inspires further neuroscience research, creating a continuous circle of progress. The insets are adapted from Rind2014aGabbiani2005aFu2018Fu2020bZhao2023Dewell2022Rind2022.
• Figure 2: The morphology of LGMD1 and LGMD2 with response to different visual stimuli. (A) The afferent network of LGMD1 (adapted from Rind2002). LGMD1 is located in the fourth layer of the optic lobe in locusts, with the retina, lamina, and medulla serving as its afferent layers. Specifically, there are $\sim 10$ neurons per lamina column with L1 and L2 proposed to be upstream of the LGMD. (B) The morphology of LGMD1 (adapted from Gabbiani2011). LGMD1 has three distinct dendritic fields where field A receives approximately $\sim 15,000$ excitatory retinotopic inputs from the entire visual hemifield, while fields B and C receive approximately $\sim 500$ non-retinotopic feedforward inhibitory inputs related to ON and OFF contrasts, respectively. Recent research has shown that dendritic field C also contributes to processing excitatory ON-contrast signals. (C) Schematic diagram of LGMD1 and LGMD2 morphology. The LGMD2 neuron (illustrated in blue) has only a single large dendrite field, and its downstream signaling pathway and postsynaptic targets remain unclear. In contrast, the downstream signaling pathway of LGMD1 has been identified as the DCMD. Dendrite field A (depicted in green) receives excitatory inputs and encodes angular velocity in the OFF contrast. The "-" symbol indicates that lateral inhibition may also occur within or preceding the dendrite of LGMD1. Dendrite fields B and C (shown in red) receive inhibitory signals from the ON and OFF pathways, respectively. Additionally, recent studies suggest that dendrite field C also processes non-retinotopic excitatory signals from the ON pathway Dewell2022. (D) The response of LGMD1 when a locust views a square projected on a screen approaching and then receding at $5 m/s$ (adapted from Rind1999). LGMD1 shows a strong and continuous spike train as the square approaches, whereas the neuron displays only a phasic spike at the onset of receding. (E) The response of LGMD1 to approaching and translating stimuli (adapted from Gabbiani2009a) - regardless of the size and speed of the translating bar, the instantaneous firing rate of LGMD1 is significantly lower when the locust views translating stimuli compared to looming stimuli. (F) The neuronal responses of LGMD1 and LGMD2 when the locust views light and dark rectangles approaching and receding, as well as in response to changes in luminance (adapted from Rind1997b) - LGMD2 is selectively excited by darker objects approaching. The long arrow indicates LGMD2 excitation after the onset of light object receding, while the short arrow indicates hyperpolarization of LGMD2 when both light and dark objects ceased during approach or at the start of receding of a dark object.
• Figure 3: Different computational models of LGMD1 and model response. (A) Illustration of a looming stimulus (adapted from Gabbiani2012). the diameter of the approaching object is $2l$, and its constant approach velocity is $v$. The angular size $\theta(t)$ projected on the retina can be calculated by $\tan(\theta) = \frac{2l}{v \cdot t}$. and the angular velocity can be determined by taking the derivative of $\theta(t)$. (B) Variation of the $\eta$-function with respect to the size-to-speed ratio $l/|v|$ (adapted from Gabbiani2011). as the ratio increases, the approach velocity decreases, leading to an earlier peak in the $\eta$-function. However, the peak neuronal response consistently occurs at a fixed delay of $\delta = 27 \text{ ms}$ after the simulated approaching object reaches a threshold angular size of $24\degree$. (C) The first four-layered LGMD1 neural network proposed by Rind et al. (adapted from Rind1999). Unlike the nonlinear mathematical model, Rind and Simmons assumed that the looming selectivity of LGMD1 is shaped by critical image cues and the critical race between excitation and inhibition within its signal processing pathway. (D) and (E) The LGMD1 network output for an approaching and receding square (adapted from Rind1999). the asterisk indicates activation of the FFI unit. The four-layered network exhibits an increasing response to looming stimuli, with the peak time of the output occurring earlier for higher approaching velocities. The network also shows a phasic response to object receding, where higher receding velocities correspond to shorter response durations.
• Figure 4: The diagram of a possible visual signal processing of LGMD from our perspective and assumption: there are three key areas represented by question marks that highlight the unknown mechanisms in LGMD circuit processing. First, it remains unclear whether a separate neuron is responsible for gathering global information and subsequently inhibiting other neurons. To date, the exact location for global inhibition is still mysterious. In line with a recent modeling work Olson2021, we assume that this inhibition is mediated by a separate neuron in the lamina layer. Second, the mechanism of lateral excitation remains poorly understood. Research has indicated that lateral excitation operates on a larger time scale compared to lateral inhibition Zhu2018, and most likely occurs within trans-medullary afferent neurons (TmAs) Rind2016. In the processing diagram presented here, we have not explicitly depicted lateral excitation since the neurons in the medulla are classified by ON/OFF contrasts, rather than as specific neurons like TmAs or DUB neurons. Moreover, we assume that ON and OFF signals at the medulla layer are completely separate, as how these signals interact is still unknown. Third, it is uncertain whether interactions occur within the dendritic trees or how integration occurs near the spike initiation zone (SIZ). In this diagram, we assume that the signals from the three dendritic trees are linearly summed without interaction, and then transmitted to the SIZ. After passing through the spike frequency adaptation, the combined signal is subsequently conveyed to the DCMD towards motor system.
• Figure 5: The hierarchical structures of three typical LGMD neural network models. (A) The four-layered LGMD network proposed by Badia et al. (image courtesy of Badia2010). This model integrates the activity of a set of ON/OFF-type neurons from the medulla that surround each location. These neurons also provide input for generating FFI. (B) The four-layered LGMD1 network proposed by Yue and Rind. (image courtesy of Yue2006a). This model introduces a artificial grouping-layer to enhance the previous LGMD network by Rind et al. Rind1996a. This improves the model's response to expanding edges while reducing sensitivity to isolated noise when coping with real-world visual stimuli. (C) The LGMD2 neural network proposed by Fu et al. (image courtesy of Fu2020). This network incorporates ON/OFF pathways and spike frequency adaptation mechanism in this framework, which significantly improve the looming selectivity. The LGMD2 neuron has only one dendrite, which shares a similar shape with dendrite field A of the LGMD1 model. Consequently, the feedforward inhibition (FFI) mechanisms present in dendrite fields B and C of the LGMD1 model are absent. To compensate for this, an adaptive inhibition mechanism is introduced to suppress the model’s response to whole-field motion. Although the frameworks presented in Figures B and C do not strictly correspond to the anatomical structure of LGMD neurons, they are widely used in LGMD-based robotic applications, supporting both ground mobile robots and UAVs for highly efficient collision-free navigation.