Robots Navigate Like Rats Using Dynamic Cognitive Maps
In a striking convergence of neuroscience and robotics, researchers have developed a brain-inspired navigation system that enables mobile robots to explore, map, and adapt to complex environments much like rodents do in the wild. Drawing on decades of research into the mammalian hippocampus—particularly the function of so-called “place cells”—a team from Beijing University of Technology and North China University of Technology has engineered a dynamic cognitive mapping model capable of real-time environmental learning, memory, and route replanning in response to unexpected obstacles.
This breakthrough, detailed in a recent study published in Acta Automatica Sinica, offers not only a novel computational architecture for autonomous navigation but also a compelling validation of long-standing hypotheses about how biological agents form internal representations of space. By emulating the neural mechanisms observed in rat experiments dating back to the mid-20th century, the team has bridged cognitive science and robotic engineering in a way that could reshape how future autonomous systems perceive and interact with their surroundings.
Place Cells and the Birth of Cognitive Maps
The foundation of this work traces back to psychologist Edward Tolman’s pioneering experiments in the 1940s. Tolman observed that rats placed in mazes could learn efficient routes to food rewards even without immediate reinforcement—a phenomenon he attributed to an internal “cognitive map” of the environment. This idea remained largely theoretical until 1971, when neuroscientists John O’Keefe and Jonathan Dostrovsky identified neurons in the rat hippocampus that fired selectively when the animal occupied specific locations. These “place cells,” as they came to be known, provided the first biological evidence for Tolman’s abstract concept.
Subsequent research revealed that ensembles of place cells form a distributed, allocentric (world-centered) representation of space, allowing animals to navigate flexibly—detouring around barriers, taking shortcuts, or recalling locations after long absences. Crucially, this system exhibits plasticity: when an environment changes, the hippocampal map reorganizes accordingly through a process known as “remapping.”
Inspired by these findings, the Beijing-based team sought to translate biological principles into a computational framework that could endow robots with similar capabilities—without relying on high-resolution geometric maps or computationally intensive simultaneous localization and mapping (SLAM) algorithms.
A Self-Organizing, Adaptive Architecture
The resulting system, dubbed the Dynamic Growing and Pruning Place Cells–based Cognitive Map Model (DGP-PCCMM), operates on a “sense–response” architecture reminiscent of self-organizing neural networks. At its core, the model uses sparse, dynamically generated nodes—each representing a “place cell”—to encode locations the robot has visited. Unlike traditional occupancy grids that discretize space into uniform cells, DGP-PCCMM creates a topological map where nodes (place cells) are added only when necessary and linked based on spatial proximity and navigability.
The process begins with a single place cell at the robot’s starting position. As the robot explores, its sensors feed positional data into the system. If the incoming data falls outside a predefined similarity threshold of existing place cells, a new cell is created. Connections between cells are established only if direct traversal is possible—ensuring the map reflects actual navigable pathways rather than mere geometric adjacency.
Critically, the model incorporates a pruning mechanism. When the robot encounters a previously open path now blocked by a new obstacle—such as a closed door—it doesn’t delete the obstructed location. Instead, it severs the connection between the current and next place cell, effectively updating the topology of the cognitive map. This mirrors the “local remapping” observed in rodent hippocampi when familiar corridors are altered.
To navigate, the robot employs a sequence planning algorithm that treats the cognitive map as a graph. Starting from its current place cell, it computes a “negative orientation function”—a proxy for path cost—toward the goal. At each step, it selects the adjacent place cell that minimizes this function, iteratively building a trajectory to the target. If an obstacle is encountered mid-route, the map is updated on the fly, and a new optimal sequence is generated without requiring full re-exploration.
Validating Against a Classic Cognitive Benchmark
To test their system, the researchers replicated Tolman’s famous detour task—a gold standard in animal cognition studies. In this experiment, a rat learns three possible routes to a food source: a short path, a medium one, and a long bypass. Initially, the animal prefers the shortest route. But when that path is blocked, it switches to the medium one. If both the short and medium paths share a common obstructed segment, the rat bypasses them entirely and takes the longest route—an indication that it understands the shared structure of the environment, not merely a sequence of turns.
The team recreated this scenario in simulation using a wheeled robot in a 130 cm × 130 cm maze with movable “doors” (simulated obstacles). With doors A and B open, the robot selected Route 1—the shortest path. When Door A was closed, it rerouted via Route 2. Most significantly, when Door B (blocking the shared segment of Routes 1 and 2) was closed, the robot immediately opted for Route 3, demonstrating an understanding that both preferred paths were compromised.
This behavior emerged without preprogramming or global path optimization. Instead, it resulted from the robot’s learned cognitive map and its ability to dynamically update connectivity. After a single encounter with a new obstacle, subsequent navigation attempts used the revised map—showing genuine one-shot learning, a hallmark of biological cognition.
Advantages Over Conventional Approaches
The researchers compared DGP-PCCMM against two widely used robotic navigation paradigms: quadtree-based adaptive occupancy grids and the dynamic window approach (DWA).
Occupancy grids, even in adaptive forms like quadtrees, require substantial memory to represent free and occupied space at multiple resolutions. In the Hairpin maze—a compact but obstacle-dense environment—quadtree mapping generated over 256 grid cells. In contrast, DGP-PCCMM used only 65 place cells to achieve full coverage, offering a more compact, semantically meaningful representation.
Moreover, occupancy grids struggle with dynamic changes. Updating them requires re-scanning and re-partitioning large portions of the map—a slow and resource-intensive process. DGP-PCCMM, by contrast, updates connectivity in constant time, making it far more responsive to transient obstacles.
DWA, a popular local planner, excels at real-time collision avoidance but lacks global awareness. In the Tolman maze trials, DWA often led robots into inefficient loops or suboptimal detours, especially in the Door B–closed scenario. Its navigation time ballooned to over 700 seconds—five times longer than in the baseline case—because it couldn’t “remember” that a previously viable shortcut was now blocked. DGP-PCCMM, having updated its cognitive map, completed the same task in just 60 seconds, demonstrating both efficiency and adaptability.
Biological Fidelity and Engineering Practicality
One of the model’s strengths is its balance between neurobiological plausibility and engineering utility. The competitive learning mechanism—where sensory input activates the “best-matching” place cell—parallels winner-takes-all dynamics in cortical circuits. The Gaussian neighborhood function used during map growth mimics lateral inhibition and synaptic plasticity observed in vivo. Even the exponential decay of learning rates over time echoes findings in hippocampal long-term potentiation.
Yet the system remains computationally lightweight. It avoids metric precision in favor of topological relationships, reducing sensor requirements and processing load. This makes it particularly suitable for resource-constrained platforms like micro-drones or exploratory robots in GPS-denied environments.
Importantly, the cognitive map is not static—it evolves through interaction. This aligns with modern views of cognition as an embodied, active process, rather than a passive reconstruction of external reality. The robot doesn’t “see” a map; it builds one through movement, memory, and correction.
Implications for Future Autonomous Systems
As autonomous vehicles, service robots, and exploration drones become more prevalent, the need for robust, adaptive spatial reasoning grows. Traditional SLAM systems, while powerful, often falter in dynamic or partially observable environments. They also demand significant computational resources and precise sensor calibration.
DGP-PCCMM offers an alternative paradigm—one grounded in biological intelligence. By prioritizing functional connectivity over geometric fidelity, it enables agents to operate effectively even with noisy or sparse sensory data. Its one-shot learning capability is especially valuable in scenarios where re-exploration is impractical, such as disaster response or planetary rovers.
Looking ahead, the researchers suggest integrating additional neural correlates—such as grid cells from the entorhinal cortex or head-direction cells—to enrich the cognitive map with metric and directional information. Future work may also explore multi-agent coordination using shared cognitive frameworks, or the incorporation of semantic labels (e.g., “door,” “staircase”) to support human-robot interaction.
A Step Toward Truly Intelligent Machines
This work represents more than a technical advance; it’s a philosophical shift. For decades, robotics has oscillated between purely reactive systems and heavily engineered world models. DGP-PCCMM occupies a middle ground—intelligent without being overdesigned, adaptive without being brittle.
By grounding robot navigation in the same principles that guide rats through mazes, the team has not only validated a century of cognitive science but also opened a path toward machines that understand space the way living creatures do: not as a grid of coordinates, but as a network of places, paths, and possibilities.
As artificial intelligence seeks ever more human-like competencies, such biomimetic approaches may prove essential—not just for efficiency, but for the kind of flexible, context-aware reasoning that defines true autonomy.
Authors: Xiao-Gang Ruan¹², Jie Chai¹², Yue Wu¹², Xiao-Ping Zhang³, Jing Huang¹²
¹ Faculty of Information Technology, Beijing University of Technology, Beijing 100124
² Beijing Key Laboratory of Computational Intelligence and Intelligent System, Beijing 100124
³ College of Electrical and Control Engineering, North China University of Technology, Beijing 100144
Journal: Acta Automatica Sinica, 2021, 47(3): 666–677
DOI: 10.16383/j.aas.c190288