Emergent Control of Ant Learning Walks in a Mobile Robot
Authors: Thomas Misiek, Andrew Philippides, Paul Graham, James Knight
Organizations: Sussex AI, School of Engineering and Informatics, University of Sussex, Brighton, UK · School of Life Sciences, University of Sussex, Brighton, UK
Before foraging, desert ants perform learning walks (LWs) to acquire visual memories around their nest. With experience, these walks expand, become straighter, and contain fewer visual scans. We previously proposed a model in which these behaviours emerge from the angular mismatch between path integration (PI) and a mushroom-body-inspired visual compass (VC). Here, we test whether this mechanism can generate learning walks on a physical robot, despite the sensory and motor noise inherent to real-world conditions. We implement the model on a four-wheeled robot equipped with a panoramic camera, an IMU and wheel encoders providing compass and stride information for PI, and test it outdoors using a replicated desert-ant experimental setup. The PI--VC mismatch regulates walking speed, turning, pirouettes, and returns to the nest, while PI provides a teaching signal for visual learning. The robot autonomously decides when to acquire visual memories, while its behaviour gradually shifts from exploration towards foraging as visual learning progresses. The robot reproduces desert-ant LW features, including expanding and straightening trajectories, decreasing pirouette frequency, and improved visual homing. These results demonstrate that a simple error-driven mechanism can generate key features of learning walks under real-world conditions, supporting its plausibility as an explanation for active visual learning in desert ants. More broadly, this work illustrates how physical robots can be used to test biological models.
Figures & tables
Fig 1: Robot platform. Photograph of the custom-built robot used in the experiments. To give an indication of scale, the wheels are 27 cm in diameter and the top-mounted panoramic camera is positioned 29 cm above the ground. The two HC-SR04 ultrasonic distance sensors used for obstacle avoidance are visible at the front of the robot, on the right side of the image. See Section 6.1 for a full description of the platform.
Fig 2: Experimental reproduction of Fleischmann et al. [12] . Photograph of the robot in the foreground, with the three white cylinders reproducing a scaled-up version of the experimental layout of Fleischmann et al. [12] , near the Library Square at the University of Sussex. The sky was cloudy. The grass is unusually yellow due to the summer 2026 drought. The buildings in the background provided prominent visual landmarks compared with the largely featureless Tunisian desert environment used by Fleischmann et al. [12] .
Fig 3: Robot point of view through the Pixpro SP360. (A) Raw 100 × 100 px fisheye image. (B) Unwrapped panoramic view of the image in A. The upper 90 ∘ , containing mostly sky, were excluded to reduce the influence of the sky and moving clouds on the visual-compass input. The three cylinders are visible but remain small relative to the surrounding landmarks such as trees and buildings. (C) Raw fisheye image acquired at the same orientation as A, but from a different position, twenty minutes after A. Distant background landmarks remain unchanged, whereas the nearby cylinders shift substantially. The clouds have also moved relative to A. (D) Unwrapped panoramic view of the image in C. Comparison of B and D also shows that the camera was not fully stabilized in pitch as the robot moved over uneven terrain. This can be seen as the yellow grassy ground is neither flat nor forming a similar horizon in B and D. Unfortunately, we did no record tilt and roll data (which the IMU had access to), so we could not plot camera pose data.
Fig 4: Top view of all 44 learning walks (A) and 54 zero-vector visual homing tests (B) conducted across the six trials. The plotted area spans 60m in longitude and 40m in latitude, with the longest LW reaching 27.32 m from the nest. Colours are used only to distinguish the trajectories and have no additional meaning.
Fig 5: (A) Example of two consecutive LWs performed by a single agent. First LW in blue and second in green. Each arrow represents the robot’s heading at 100 ms intervals. Red arrows indicate the robot’s heading during snapshots. Nest position as a white star. Cylinders as white circles. The background is a 50% whitened Google Earth image of the experimental site. (B) Example of two consecutive ZV visual homing trajectories from an experienced agent. Seventh test in blue, eighth in green. Tests started 12 meters south-east of the nest (red dot mark). Both visual searches are centered roughly on the nest. Scale differs between A and B.
Fig 6: All consecutive LWs of two agents. (A) All consecutive LWs for agent 2. Eight walks were recorded, each shown in a different color (blue for the first walk, orange for the second, etc.). The 6 m radius around the nest is indicated by a grey dashed circle. The first two LWs remain close to the nest. Subsequent walks are more directed and extend beyond the 6 m radius, and are therefore labelled as foraging walks. The plotted area is a 24×24m square. (B) Another example of a series of consecutive LWs, for agent 4. Nine walks were recorded. Cylinders as white dots.
Fig 7: Changes in LW behaviour with increasing experience. Measurements are shown for the first four LWs, with all subsequent walks grouped as 5+. (A) Maximum distance from the nest increased approximately threefold between the first LW and the 5+ group. (B) Total distance travelled increased across LWs. (C) LW duration remained broadly similar, with a slight increase across walks. (D) Path straightness doubled between the first LW and the 5+ group. (E) Pirouette frequency decreased two-fold. (F) Number of pirouettes performed within a 6 m radius of the nest, following the same criterion [ 33 ] , decreased twofold. (G) PI–VC angular error, which control learning behaviours, decreased across LWs (H) Proportion of foraging walks ( >6 m from the nest), increased and reached a plateau by the third LW. (I) Forward driving speed of the robot during successive LWs in meters per seconds.
Fig 8: Visual homing tests performance at different levels of experience. (A) Heatmap of all naive test trajectories. Cylinders as white dots, nest position as a white start. Grid size is 4 m by cell. Test start position as a red dot. ( n=6 ) (B) Heatmap of all homing tests from long-LW agents. Searches remain spread out, but are clearly centred on the nest. ( n=23 ) (C) Heatmap of all tests from foraging agents ( n=25 ). Trajectories are more directed and searches are more tightly centred on the nest. Long LWs data not shown. (D) Centroid distance to the nest during visual homing tests as a function of test ID. Search centroids were farther from the nest during early tests and became progressively closer after successive LWs, showing improved homing performance with experience.
Fig 9: Influence of the cylinders (A) Search behaviour was affected by the position of the cylinders. The agent was trained using cylinders positions shown in blue. Control tests shown in blue. All tests were conducted with an experienced agent which performed seven LWs prior to the tests. The orange trajectory shows the search after the cylinders were shifted 4 m to the northwest. The orange, shifted configuration was used only during tests. The search centre shifted in the same direction as the cylinders, showing that they strongly influenced visual homing. This is reminiscent of the classic landmark-displacement experiment of Tinbergen [42] , in which shifting landmarks around a wasp nest displaced the subsequent search. (B) Nevertheless, the cylinders were not the only landmarks available to the agent. When they were removed, the agent was still able to perform a tight search centred near the true nest position, indicating that other visual landmarks, such as trees and buildings, could also support homing.
Fig 10: Pirouette and snapshot statistics. (A) PI angular error relative to the ground-truth nest direction provided by GNSS. (B) Distribution of snapshot headings relative to North, showing an approximately uniform distribution. (C) Distribution of snapshot headings relative to the nest direction, also approximately uniform. (D) Angular interval between consecutive snapshots, with two peaks around −65∘ and +65∘ . (E) Distance of snapshots from the nest. Most snapshots, and therefore most pirouettes, occurred within 5 m of the nest. (F) Inter-pirouette time interval averaged across LW1 and LW2, with a peak around 7 s.
Fig 11: Comparison of scanning-bout timing in ants and robotic learning walks. (A, B) Distribution of interbout intervals during the first (A) and second (B) learning walks of ants, digitized from Deeti and Cheng [8] , with mean intervals of approximately 2.15 and 2.45 s, respectively. (C) Mean time of successive scanning bouts in ants, with linear regression slopes of 2.16 and 3.87 s/bout for LW1 and LW2, respectively ( R2=0.98 and 0.99 ). (D, E) Distribution of inter-pirouette intervals during the first (D) and second (E) learning walks of six robotic agents, with mean intervals of 6.31±1.01 s ( n=70 ) and 7.16±1.35 s ( n=71 ), respectively (mean ± SD). (F) Mean time of successive robotic pirouettes, with regression slopes of 6.29 and 7.13 s/bout for LW1 and LW2, respectively ( R2=0.9998 and 0.9995 ). Error bars in (C) and (F) represent SD. Data in (A–C) were approximately digitized from Deeti and Cheng [8] .
Fig 12: PI search error after LWs. (A) PI search locations after the 44 LWs. Red dots indicate the positions at which the PI search started, the black star indicates the true nest position, and the black cross indicates the centroid of the search locations. Grey lines show the trajectories produced when the robot was manually driven from the PI search position back to the nest. The background is a 50% whitened Google Earth image of the experimental site. The plotted area is 16×16 m. (B) Percentage of snapshots for which the angular error between the PI-estimated nest direction and the GNSS-derived ground-truth exceeded 90∘ , for four distance bins from the nest. (C) Absolute angular error between the PI-estimated and GNSS-derived nest directions for the same distance bins. Only measurements with GNSS horizontal accuracy below 10 cm were included in B and C. Average angular error decreased with higher distance to the nest.
Fig 13: Schema of the robot platform
Fig 14: GNSS horizontal accuracy. (A) Horizontal accuracy over time for agents 1–4, tested on 15 August 2026. Each line corresponds to a LW (blue) or visual homing test (red). Most measurements were below 5 cm and nearly all below 20 cm; the worst was approximately 2 m. (B) Distribution of horizontal accuracy across the four GNSS statuses for agents 1–6. Blue (left flattened box): RTK fixed; orange: RTK float; green: no RTK; red: bad GNSS fix. Accuracy was high for RTK fixed and RTK float states, ranging from very precise in RTK fixed ( 1.4 cm) to unusable in bad GNSS fix (¿10m). (C) Horizontal accuracy for agents 5 and 6, tested on 16 August 2026. Accuracy exceeded 2 m in four walks and even reached up to 33 m. One walk showed sustained poor accuracy above 2 m. (D) Proportion of recording time spent in each GNSS status. Agents 1–4 (left), tested on 15 August, spent almost all recording time in RTK fixed or RTK float states. Agents 5–6 (right), tested on 16 August, spent more time without RTK and nearly 1% of the data was completely unusable due to bad GNSS fix.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Fig 15: Relationship between prediction variance and visual-compass accuracy on the robot. (A) Absolute angular error of the centroid of 10 consecutive raw visual-compass predictions (1 s of sampling), relative to the ground-truth nest direction derived from GNSS position. During these 1-s windows, the maximum angular spacing between headings was 54.31∘±20.63∘ (mean ± SD), with a maximum of approximately 100∘ , while the maximum linear spacing between sampled positions was 0.287±0.212 m, with a maximum of approximately 0.8 m. Only LW3 and subsequent LWs were included, excluding LW1 and LW2. Mean centroid error increased monotonically with circular variance, from 20.42∘ in the lowest-variance bin ( 0 – 0.1 ) to 83.32∘ in the highest-variance bin ( 0.9 – 1.0 ). Prediction variance was positively correlated with centroid error (Spearman’s ρ=0.358 , p<0.001 ), with a linear R2=0.167 and mutual information of 0.113 nats. Error bars show standard deviations, which increased with variance. The variability in all variance bin was quite high, hence the low correlation and low predictability despite a clear trend on average. (B) Frequency distribution across all the one-second prediction samples across circular-variance bins. Low-variance samples were more frequent. The positive correlation between prediction variance and angular error suggests that the variance of successive visual-compass predictions could serve as a proxy for prediction accuracy, providing a confidence or reliability signal. This could be used to weight visual-compass predictions against PI, or to distinguish familiar from unfamiliar locations. Unlike the visual familiarity or novelty signals commonly derived from MBON activation in visual-compass models [ 34 , 41 , 1 , 48 ] , this signal would reflect the consistency of successive directional predictions rather than the familiarity of the visual input itself.
Quadruped locomotion provides a natural setting for understanding when model-free learning can outperform model-based control design, by exploiting data patterns to bypass the difficulty of optimizing over discrete contacts and the combinatorial explosion of mode changes. We give a principled analysis of why imitation learning with quadrupeds can be inherently effective in a small data regime, based on the structure of its limit cycles, Poincaré return maps, and local numerical properties of neural networks. The understanding motivates a new imitation learning method that regulates the alignment between variations in a latent space and those over the output actions. Hardware experiments confirm that a few seconds of demonstration is sufficient to train various locomotion policies from scratch entirely offline with reasonable robustness.
Insect neuroethology provides a compelling biological template for efficient autonomous navigation. We draw an analogy between the formal embodied AI visual 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. We develop a novel integrative model of mushroom body and central complex, two insect brain structures, that have been implicated, respectively, in associative learning and path integration. We demonstrate the mushroom body learning triggered by collisions results in adaptive obstacle avoidance and consequently optimised paths to the goal, corroborating the hypothesis of recent behavioural work that an insect can learn continuously as they travel. The embodied insect-inspired model achieves success rates comparable to recent state-of-the-art models at many orders of magnitude less computational cost in the standardised Habitat point-goal navigation benchmark. Testing in a more realistic simulated environment validates its robustness to perturbations.
Yihe Lu, Barbara Webb
School of Informatics, University of Edinburgh, Edinburgh, Scotland, United Kingdom · University of Glasgow, Glasgow, Scotland, United Kingdom
Legged robots with egocentric forward-facing depth cameras can couple exteroception and proprioception to achieve robust forward agility on complex terrain. When these robots walk backward, the forward-only field of view provides no preview. Purely proprioceptive controllers can remain stable on moderate ground when moving backward but cannot fully exploit the robot's capabilities on complex terrain and must collide with obstacles. We present Look Forward to Walk Backward (LF2WB), an efficient terrain-memory locomotion framework that uses forward egocentric depth and proprioception to write a compact associative memory during forward motion and to retrieve it for collision-free backward locomotion without rearward vision. The memory backbone employs a delta-rule selective update that softly removes then writes the memory state along the active subspace. Training uses hardware-efficient parallel computation, and deployment runs recurrent, constant-time per-step inference with a constant-size state, making the approach suitable for onboard processors on low-cost robots. Experiments in both simulations and real-world scenarios demonstrate the effectiveness of our method, improving backward agility across complex terrains under limited sensing.
Shixin Luo, Songbo Li, Yuan Hao +4
The authors are with Institute of Cyber-Systems and Control, Zhejiang University, 310027, China. · The author is with College of Artificial Intelligence, Nankai University, 300350, China. · The author is with Hangzhou Public Library, 310016, China.