1. Introduction

Survival in complex environments requires efficient navigational strategies. From desert ants to humans, successful wayfinding, defined as navigating toward goals that are not directly visible, depends on emergent internal spatial representations, known as cognitive maps [1,2]. Understanding how such maps are constructed from ongoing experiences and how they can be exploited for flexible goal-directed navigation remains an active area of research in both neuroscience and reinforcement learning.

In the brain, the hippocampus (HP) and the entorhinal cortex (EC) are the central areas involved in spatial representation. They contain specialized neurons encoding spatial and contextual information, including grid, border, speed, and place cells [3–5]. In particular, the latter are numerous in the CA1 hippocampal sub-region and have attracted interest due to the convergence of inputs from entorhinal grid cells, the CA3 hippocampal sub-region, and the lateral EC [6–9]. This strategic integration of diverse spatial and contextual signals suggests that CA1 place cells may play a critical role in the formation and maintenance of cognitive maps [10].

Traditional theories of cognitive maps suggest that spatial representations can emerge from multiple navigational strategies. Early frameworks suggested that the hippocampus encodes both spatial location and direction [11], while graph-based models captured structural aspects of spatial organization, although often at the cost of scalability [12].

Another approach is path integration, which involves tracking one’s position by integrating past movements—using actions or velocity vectors—to estimate the trajectory and return to the point of origin.

An important strength of path integration lies in its independence from external cues, relying instead on idiothetic internal-motion signals, which are considered biologically plausible and believed to involve entorhinal grid cells [13,14]. A common computational methodology to investigate path integration involves recurrent neural networks combined with linear readouts. When trained in navigation tasks, these systems spontaneously develop spatially tuned activity patterns reminiscent of grid, place, and border cells [15–17].

Another proposal is the Tolman-Eichenbaum machine (TEM), a model that generalizes across spatial and relational tasks while reproducing key biological neural activity patterns [18–20]. Learning in TEM occurs through gradient descent over many episodes or epochs, in contrast to adaptive, context-dependent learning observed in biological systems. As with several modeling approaches, its neuronal dynamics remain simplified, typically reduced to artificial neurons defined by weighted sums and fixed non-linear activation functions. An alternative is the successor representation framework (SR), built around the learning of a predictive map of the explored state space, which can be used for active navigation [21,22]. It has been linked to hippocampal cognitive map generation, with the possibility of incorporating reward information and formulation in terms of biologically plausible mechanisms [23–26].

Neuromodulation is another relevant element in brain dynamics, affecting physiology, cognition, and behavior [27]. Its functions include filtering meaningful internal and external signals, influencing neuronal dynamics, and learning by altering synaptic states and parameters [28–30]. An important neuromodulator is dopamine, long associated with reward information, prediction error encoding [30,31], and novelty detection [29], mechanisms closely related to reinforcement learning principles [32,33]. In addition, projections from the Ventral Tegmental Area (VTA) and Locus Coeruleus (LC) have been shown to target the hippocampal circuits [34–36], to be involved in memory formation [37], and affect the adjustment of CA1 place cells [38–41], particularly through LEC inputs [42,43].

Other approaches have incorporated neuromodulators with reward-driven Hebbian plasticity [44] or other learning features, such as sparsity and update scaling [45]. However, these architectures often remain limited in their ability to integrate these components into a biologically grounded framework that works without extensive training, gradient-based optimization, or external coordinate inputs while still providing adaptable planning and behavior.

In this work, we addressed these limitations by proposing a bio-inspired model of cognitive map formation learned online from sensory experiences. The model builds on the emergence of place cell representations from grid cells activity and velocity information, that is, internally available idiothetic input, in line with the path integration approach [15,16]. We designed plasticity mechanisms through which neuromodulatory signals could target and modulate place cells, augmenting the resulting map with behaviorally relevant information. Combined with a network-based path planning mechanism, our architecture enabled an artificial agent to construct content-rich topological maps of its environment on-the-fly and to leverage them for targeted navigation. We took direct inspiration from previous work discussing the involvement of neuromodulation in hippocampal spatial processing, for example, in terms of the gating of incoming input streams and the influence of synaptic plasticity [46–48].

Our primary aim was to evaluate the potential of the emergent cognitive map to support navigation in unexplored environments, including the ability to record the location of a reward and to generate paths to it from already visited regions. Notably, the model begins without any spatial or qualitative knowledge of its surroundings; such knowledge is acquired incrementally as the agent moves. The lack of separation between map construction and reward collection ensures that the task operates entirely online.

A second aim was to assess the model’s robustness and adaptability to qualitative and spatial changes, such as shifts in reward location or the introduction of new boundaries. Additionally, the contribution of neuromodulation to the regulation of place cell activity was also examined as experience accumulated, investigating the effect on cell density and the size of place fields. Finally, we sought to quantify the effect of neuromodulation on task performance through a series of comparative ablation experiments.

These analyses allowed us to characterize the effects of neuromodulation on place field properties, such as modulation of their size and density, and to discuss potential links to experimental findings [49,50].

The remainder of the paper is organized as follows: Section 2 details the model and experimental setup; Section 3 presents results; Section 4 discusses the broader implications and future directions.

2. Methods

We propose a model of cognitive map formation driven by an agent’s experience within a closed environment.

The architecture operates with minimal external inputs, limited to binary reward and collision signals, as illustrated in Fig 1A. Instead of relying on exteroceptive cues, spatial representations emerge from idiothetic information, that is, the agent’s internal perception of self-motion [51], consistent with path-integration frameworks. In practice, we use the agent’s ground truth velocity vector, that is, its actual displacement within the environment, as the primary navigational signal, reflecting the integration of inertial and proprioceptive signals observed in biological systems [13,52]. Since no visual information is used, the agent effectively navigates in the dark. Importantly, the model does not receive allocentric coordinates as sensory input; any two-dimensional positions used for planning or visualization are derived from the internally constructed place cell representation.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Model layout and spatial representations.

a: layout of the architecture consisting of four main inputs: a reward signal targeting the dopamine modulation component (DA), a collision signal targeting the boundary modulation component (BND), a velocity vector (the actual movement made by the agent) aiming at the grid cells module, and a reward trigger acting as a flag to enable reward-directed navigation. A place cells component represents where a cognitive map actually resides, and receives input from the grid cells for spatial information, modulators for non-spatial data, and it is used for navigation and determining the reward position. Two policy components are responsible for determining the agent’s action: an explorative behaviour and a reward-directed behaviour, the latter being conditioned on the ability to calculate the reward position and the presence of a trigger signal. b: visualization of part of the path-finding algorithm, propagation of an activity wave through the place cell network from top-left to bottom-center, and the calculated path visualization in the bottom-right. c: a module of grid cells defined in a bounded square space of length 1, and an activity representation of their receptive field over a torus. d: formation of a cognitive map during navigation, with place cells (grey) being formed over the agent’s trajectory (black line); boundary-modulated place cells are formed near boundaries (blue). e: spatial distribution of the place cell centers, together with the place fields of two cells. f: representation of the cognitive map as place cells tagged by neuromodulators during collisions (blue), reward events (green), or untagged (grey); the color gradient correlates with the intensity of the neuromodulators’ weights. g: illustration of a path (arrows) over a place cell map connecting an initial position (red) and a goal (green).

https://doi.org/10.1371/journal.pcbi.1013487.g001

2.2. Neuromodulation

Neuromodulators deliver event signals: rewards, denoted DA (for dopamine), and boundary collisions, denoted BND. In particular, the latter is an internal variable that correlates with and tracks the sensory experience of mechanical collisions.

They are driven by binary inputs and defined through a leaky variable with exponential decay.

Each modulator k updates its connection weights with the place cells through a Hebbian rule based on place cell activity and the error:

(1)

The term in brackets can be interpreted as an error term that implements a simple form of predictive coding, and is inspired by temporal-difference learning [33], aligning with evidence that neuromodulatory systems signal prediction errors and update beliefs [19,55,56]. This mechanism was intended to support robustness to environmental changes, such as reward displacement or the introduction of new obstacles.

The weight vectors are restricted to remain non-negative. Reward modulation tags cells near reward locations, whereas boundary modulation builds a representation of the environmental edges. The result is a set of scalar fields over the place cells, and it forms the core of our model of a cognitive map (Fig 1F). Throughout the manuscript, the architecture is understood as containing only place cells: labels such as reward-modulated (or DA-tagged) and boundary-modulated (or BND-tagged) refer to subsets of place cells carrying non-zero modulatory weights, not to distinct neuron types. See the Supplementary Information for full learning rules and parameter settings.

2.3. Modulation of place fields

In addition, we tested whether neuromodulators could directly alter spatial tuning. The place fields were dynamically shifted and resized according to recent salient events, effectively changing the cell density and the area of the receptive fields.

More in detail, following a reward or collision signal, the place field centers were displaced within the grid-cell space according to a vector whose magnitude depended on the neuromodulator and the spatial proximity to the reward or collision location:

(2)

Here, is a Gaussian function with width and a scaling factor. This rule is inspired by BTSP plasticity [49], which shifts CA1 place fields following salient experiences. This action was applied only to recently active cells, that is, with an activity trace greater than a given threshold . The motivation was to prioritize activity within a recent time window, consistent with suggestions of dopaminergic gating in the hippocampus [57,58]. Furthermore, this approach is consistent with mechanisms such as the BCM rule [59] and the aforementioned BTSP, both of which rely on asymmetric kernels that reduce or filter out inactive neurons [39].

Lateral inhibition prevents field overlap during remapping. Moreover, field size is modulated by scaling the gain of recently active neurons, allowing neuromodulators to transiently enhance or suppress the spatial sensitivity of specific cells. This modulation rule involves the gain of each cell being adjusted proportionally to its activity trace , a reference gain constant , and a modulatory scaling variable.

(3)

where is a scaling gain parameter, for which a value of 1 means absence of modulation.

Finally, these constitute the mechanisms through which place fields could be reshaped, and the cells that were not affected maintained stable and unchanged fields.

3. Results

The main objective of this work was to assess the model’s ability to construct an adaptive spatial representation that supports effective goal-directed behavior.

To evaluate the model, we generated a set of closed environments with an increasing number of internal walls, thereby varying the difficulty of mapping and navigation. The internal arrangement of walls and corners renders these environments non-convex. They thus qualify as a wayfinding setting [63], where the goal location is hidden from the agent’s immediate line of sight. At each time step, the agent’s selected action updated its position. A binary collision signal was generated (1 if the new coordinates intersected or crossed a wall, and 0 otherwise). Similarly, the reward signal was defined relative to the goal location, delivering a value of 1 with a probability of when the agent entered a circular region of radius around the reward, and 0 otherwise.

A simulation trial was defined by initializing the model without any place cell or learned weights and placing the agent randomly. To facilitate map analysis by increasing the average number of place cells formed, the reward position was initially hidden, becoming available only after a fixed duration. Once available, each time a reward was collected, the agent was randomly teleported to a new location; concurrently, the model’s grid cell modules were re-calibrated to keep the internal map and newly recruited place cells aligned.

We defined the performance metric within a trial as the total number of rewards obtained over a duration T_dur.

This evaluation of map formation and performance within the same trial reflects the online nature of the model, in contrast to approaches based on distinct multi-epoch training and testing stages.

Following these simulation settings and parameter optimization, the remainder of this section is organized around three main questions. First, we investigate whether the model can form a cognitive map online to navigate a representative environment (Figs 2A and 2B). Second, we illustrate how this map is locally refined and updated during specific tasks, including place field modulation, detour behavior, and reward relocation (Figs 2C and 2H). Finally, we quantify these event-dependent changes and establish the functional role of neuromodulation via ablation studies (Figs 3A and 3E).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Cognitive maps and performance results.

a: illustration of a cognitive map composed of boundary-modulated place cells (blue), reward-modulated place cells (green), and a line (red) representing the planned trajectory to reach a target location from the current position. b: representation of the same environment as in panel a but with the reward area shown as a circle (green), together with the past trajectories (red line) and the agent position (black square). - c: a cognitive map in a different environment showing shifts of the place fields for the two types of modulated place cells. Specifically, the starting boundary-modulated place cell centers (magenta circles), final boundary-modulated place cell centers (blue circles), and their displacement vectors with length proportional to the magnitude (blue arrows) are shown. Similarly, they are shown the starting reward-modulated place cell centers (orange circles), final reward-modulated place cell centers (green), and their displacement vectors (green arrows). Finally, there are also non-modulated place cells (grey filled circles). - d: example cognitive map representing place field size modulation. Both reward-modulated place cells (green) and boundary-modulated place cells (blue) are shown as circles whose radius represents the relative width of each cell’s place field, calculated from its gain parameter : larger values correspond to smaller fields. This plot is intended as a qualitative visualization; the corresponding statistical comparison across simulations is reported in Fig 3C. - e: detour experiment, plot of trajectories before (black) and after (red) the insertion of a wall (rectangle) between the starting and goal positions; the wall is also indicated by the boundary-modulated place cells in blue. - f: collision (blue) and reward (green) counts during a simulation trial where a new wall was inserted after 5000 time steps (red dotted line). Curves represent a 1000-step moving average across 64 repetitions, with shaded regions denoting the standard deviation. - g: trajectories across multiple trials in an example environment for the reward-relocation task, with the agent starting at the same position (black square) and the reward location periodically moved among the green circles. - h: same legend and protocol as in plot f but with the reward location changed instead.

https://doi.org/10.1371/journal.pcbi.1013487.g002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Statistics and performance results.

The labels “BND cells” and “DA cells” abbreviate boundary- and reward-modulated place cells, respectively. a: effect of collision count on the number of BND- and DA-modulated place cells (top) and on their average activation gain (bottom). Each sub-plot reports Pearson’s r, R², and the p-value. - b: same layout as in plot a but using reward count. - c: Dunnett-corrected differences in place field center displacement (top) and average gain magnitude (bottom) for DA-modulated, BND-modulated, and non-modulated control place cells; larger values correspond to narrower place fields. - d: reward-count performance across the chance, no-modulation, full-modulation, DA-only, BND-only, density-only, and gain-only variants. Significant pairwise differences are color-coded as: grey <0.05, black <0.01, and red < 0.001. Results obtained from pairwise t-tests across 128 runs with Bonferroni correction; the numerical report is provided in S4 and S5 Tables - e: same analysis as plot d but for collision count, calculated from the time a reward representation is initially formed (set at the first non-zero DA weights); the collision count of the control model was not considered, since it lacked reward-driven navigation. The numerical results are reported in S6 and S7 Tables.

https://doi.org/10.1371/journal.pcbi.1013487.g003

4. Discussion

Exploration and planning in novel and past environments are essential behaviors of animals that directly affect their success in spatial understanding and achieving goals.

An important element behind these abilities is the formation of a map of their surroundings, enriched with information gained from new experiences, known as a cognitive map. In this work, we presented a rate network model that simulates the CA1 area of the hippocampus [10,70]. It differs from previous approaches as it operates online, uses neuromodulated synaptic plasticity, and does not rely on external coordinates.

We used simplified grid cells and synaptic plasticity to generate information-rich spatial representations rapidly forming in novel environments, dynamics directly following experimental observations and computational modeling of place cell generation [71,72]. The architecture and its mechanics are based on the proposal that CA1 place coding is supported by entorhinal grid cell modules that operate at different spatial frequencies [73,74]. In the spirit of minimizing geometric assumptions in the neural space, we treated the generated place network as a topological graph learned through velocity inputs, reminiscent of path integration [14,75,76]. External sensory information was added locally to each graph node, i.e., a place cell, through the action of neuromodulators acting as a signal filter of salient events such as boundary collisions and reward occurrences. We took inspiration from experimental findings about the involvement of dopaminergic projections in hippocampal representation [65,77]. This idea of enriching a spatial map aligns with the concept of a labeled graph [78,79], a cognitive map proposal that addresses the problems of geometric and metric consistency by harnessing noisy observations and supporting vector-like operations [80–82].

Although incorporating a richer set of sensory inputs, such as additional self-motion cues, could further improve performance, our goal in this work was to provide a proof of concept. Specifically, our interest was to document the emergence of advantageous computational strategies even from relatively simple modeling assumptions.

The tasks we applied the agent to consisted of an exploratory and exploitative phase, in which the system was tasked to plan and reach reward positions. For simplicity, the first stage relied on a random walk process, as more elaborate policies were outside the scope of this work. This choice had the side effect that the reward was not always discovered, leading to the formation of incomplete maps and impairing performance. However, this constraint did not qualitatively alter our central findings.

The simulation results support the model’s central assumptions, showing the formation of cognitive maps and their encoding of information collected during experience.

Previous work has used path integration with deep neural networks, but required extensive gradient-based training [13,16,17]. Another important difference is that our resulting neural network was composed solely of place cells, although neuromodulated, and no other types of neurons were present. This distinction is justified by the partially different task protocol and internal architecture, which constrained the tuning dynamics and also did not receive visual information as in [15]. Furthermore, our model relied on predefined grid cell layers, which constituted a strong and sufficient inductive bias, and did not have to be learned from scratch. The selection of the goal location was based on dopaminergic connections and a spatial average of the associated neurons. This approach is consistent with experimental observations and dopamine activity recordings in CA1, suggesting its role in stabilizing long-term spatial memories and supporting reward-guided navigation [60,65,83].

An additional relevant aspect is the consideration of the place cell layer as an explicit graph data structure, on which path-planning was applied, meant to be a neural-based variant of the Dijkstra algorithm [84], a popular choice for graph computations. In fact, our algorithm relied on biologically plausible neural operations, such as activity propagation through the network and the use of synaptic activity traces as masks to identify the best neighboring place cells. This structured representation introduced a strong inductive bias, promoting robustness and flexibility across environments with varying layout complexity. Moreover, this lifted the need to learn an approximation of it through network dynamics and rely on a greater variability in neural receptive fields.

A similar approach using boundary-modulated place cells has also been used for applying the successor representation (SR) framework to large environments [85], with the emergence of place and grid cell tuning. More broadly, on the one hand, several aspects of the SR framework align with our model: the distinction of the state map and the reward/boundary signals, and the use of an update rule inspired by temporal difference learning [21–23]. In addition, several studies have investigated biologically plausible implementations [24–26], obtaining close and promising matches with experimental results. On the other hand, the construction principles of our architecture required fewer assumptions about the state space size (the environment area can be almost arbitrarily large), while learning the basis of the state representation (place cells) and the successor matrix (associated with the lateral connectivity) is more direct and immediate than in the standard SR formulation. These modeling choices reflect our interest in the online generation of an explicit place cell map with an associated graph representation, while the SR may also produce non-spatial neural profiles. In fact, the type of cognitive map that emerged in our results was solely composed of neurons with a well-defined tuning curve requiring a single grid cell population vector, thus not demanding multiple spatial observations or learning steps, except for adapting the modulation weights to failed predictions. Finally, our architecture was better suited to integrate and study our formulation of neuromodulated plasticity, given direct access to a spatial map.

Adaptability was tested by occasionally moving the reward position, leading to the generation of an internal prediction error that was used to update its representation on the map. The agent was able to weaken previous associations, return to exploration, and encode new reward locations. This behavioral protocol is similar to previous work [86], in which dopaminergic and cholinergic activity was utilized within a Hebbian plasticity rule to strengthen or weaken reward-associated spatial representations. However, as an alternative to exploiting neuromodulators with opposite valence, we followed the direction of temporal-difference learning and predictive coding, a framework linked to hippocampal representations [19,87] and explored by various computational approaches [88–90]. This preference arose from our focus on using the data encoded in the cognitive map itself, in which each position has an associated neuromodulation value array, encapsulated in the modulatory connection weights. This array functions as an expected sensory experience to compare with the actual sensory experience to obtain a learning signal. This mechanism is aligned with long-standing theories linking neuromodulation, particularly dopamine, with prediction error signaling [31,38,66,91,92]. It also fits within the larger framework of predictive coding, which interprets cognition as the minimization of surprise [90,93], and matches with temporal-difference learning, where the values of internal states are continuously updated through sensory feedback [19,21,33].

However, a limitation of our current implementation is the strong simplification of neuromodulation to externally defined sensory events represented as a binary two-dimensional input.

An important consequence of this choice is the marked limitation of the information about the environment available to the agent, which is effectively restricted to boolean quantities. This operationalization differs substantially from biology, where signals are encoded in a distributed manner and integrated with multiple pathways, providing neuromodulation circuits with high-dimensional neural representations and richer responses [47,57]. Furthermore, neuromodulation in our model only influences value representation and place field modulation, which captures only a subset of the larger roles neuromodulators play in the brain [34].

Lastly, the relevance of active modulation of the neuronal properties of place cells was supported by simulated ablation experiments. The primary finding was the statistically significant role of active boundary modulation, obtained by comparing it with a model variant with disabled neuromodulation. This supports the importance of having a flexible representation of the environment layout.

Further investigation involved testing the effect of density and gain modulation separately. The results showed that altering the density of the place cells affected the total count of the rewards collected. These results are consistent with experimental observations that demonstrate the remapping of place fields following salient events [68,94]. Notably, behavioral time-scale plasticity (BTSP) has been shown to induce shifts in CA1 place fields in response to salient inputs delivered through pathways from the entorhinal cortex [39,49,72,95]. Additional studies further support this view, reporting spatial reorganization of place cells to better emphasize behaviorally relevant information [96,97], as well as changes in firing rate following contextual stimuli [98,99], or other activity processes [100].

The effect was particularly clear for boundary-modulated place cells, which displayed a trend of shifting place cell centers closer to the walls, yielding a statistically significant improvement in performance. One possible interpretation is that a tighter representation of environmental edges leaves more room for the interior, improving its spatial resolution, and thereby enhancing mapping and supporting more efficient path planning. Consistent with this perspective, previous studies have reported an increase in the clustering of place field edges near surface boundaries [101], and have shown that the presence of visual cue edges stabilizes spatial tuning [102]. It should be noted that in our model, the boundary information was limited solely to collision signals rather than feature-rich sensory input. Nevertheless, the emergence of strategies that parallel biological observations suggests the existence of shared computational advantages, supporting previous work documenting the benefit of increased cell clustering near boundaries for spatial accuracy and context discrimination [69,103].

Regarding reward-modulated place cells, despite the reliable experimental evidence to support their remapping, the effect was weak and did not substantially improve behavior, a pattern that likely followed from the baseline constraints of our simulation settings.

Concerning the modulation of place fields, there is significant experimental evidence of their alteration during reward events [104–106], some reporting shrinkage near reward objects [107], and more markedly near boundaries [108]. In this direction, two possible experimental predictions can be made, based on the constraint of using self-motion cues for the generation of cognitive maps.

First, in the model, place field sizes near boundaries are initially approximately homogeneous, and are reshaped only after meaningful events, such as collisions. This can be tested by correlating individual place-field structural metrics with the cumulative count of relevant events occurring near their fields. Secondly, a similar test could be performed for rewarding events, potentially going further by blocking dopamine release in order to assess its causal involvement.

In our setup, the place-field modulation was implemented by scaling the neural activation gain. This led to a general reduction in field size across all modulated cells.

In the case of boundary-modulated place cells, this shrinking effect may account for the improved performance in two ways: first, by increasing the spatial resolution of the environment layout [109]; and second, by freeing up space for other place cells to form nearby. The latter may result from reduced lateral inhibition and the inward shift of the field center toward the boundaries.

However, gain modulation alone did not produce statistically significant effects, suggesting that, in this protocol, it may need to be coupled with density modulation to produce measurable improvements.

For reward place cells, the impact of gain modulation was less pronounced. This may be due to the limits of our experimental protocol or reward signal, with the latter being defined purely in spatial terms and lacking any meaningful non-spatial feature.

Previous studies have proposed that the Locus Coeruleus (LC) is involved in place cell reorganization [41] and overrepresentation near reward locations through the co-release of dopamine and noradrenaline [40]. A possible experimental prediction that emerges from our results is that preventing place cell density modulation may indirectly decrease the quality of reward-directed navigation, as suggested by the remapping of boundary-modulated cells shown in Fig 2C and the statistics reported in Fig 3D.

More specifically, this can be tested in animal models by pharmacological blockade of LC afferent connections to CA1, once a reward has been experienced. Then, performance can be evaluated by considering: the count of rewards fetched; the difference between the animal’s trajectory between the initial and reward location and the theoretical spatial geodesic (shortest path); and the average distance from the walls. However, a relevant confounding variable in this setup is the difficulty of isolating CA1 inputs, alongside the involvement of multiple and complex additional signals in the generation of the cognitive map.

In conclusion, this work showed a possible architecture for coupling emergent spatial representations with neuromodulated plasticity to achieve an experience-driven cognitive map. The reliance on some spatial and algorithmic inductive biases, grid cells, and a planning algorithm supports the idea of a labeled graph for goal navigation. Future work can investigate the application to other spatial domains, such as motor control and three-dimensional navigation. In addition, a richer input feature can be added, such as visual information [110], as well as new neuromodulators that encode different sensory dimensions or internally generated signals.

Supporting information