Collective Evolution Learning Model for Vision-Based Collective Motion with Collision Avoidance
Abstract
Collective motion (CM) takes many forms in nature; schools of fish, flocks of birds, and swarms of locusts to name a few. Commonly, during CM the individuals of the group avoid collisions. These CM and collision avoidance (CA) behaviors are based on input from the environment such as smell, air pressure, and vision, all of which are processed by the individual and defined action. In this work, a novel vision-based CM with CA model (i.e., VCMCA) simulating the collective evolution learning process is proposed. In this setting, a learning agent obtains a visual signal about its environment, and throughout trial-and-error over multiple attempts, the individual learns to perform a local CM with CA which emerges into a global CM with CA dynamics. The proposed algorithm was evaluated in the case of locusts swarms, showing the evolution of these behaviors in a swarm from the learning process of the individual in the swarm. Thus, this work proposes a biologically-inspired learning process to obtain multi-agent multi-objective dynamics. Author summaryMulti-agent multi-objective tasks are common in nature with examples as collective movement in birds and economic management in humans. These problems are famous for being convoluted to efficiently solve. Nonetheless, nature has been successfully solving it for millennials using an evolution strategy. A prominent example is a task of flocking performed by multiple species, which involves both collective motion and collision avoidance. In our work, we simulate agents that are able to learn behaviors on the individual level, that in turn translate into the desired group (multi-agent) behavior. Using nature-inspired genetic algorithms and reinforcement-learning methods, the agents are successfully implementing a collective behavior similar to the one encountered in nature.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 19
- FWCI
- 2.11
- Field percentile
- top 13%
- Organism
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 3 (12 Mar 2023) · CC BY 4.0
Q-learning policy trained on genetic-algorithm targets lets simulated locust-shaped agents with only visual input flock (CM score 0.7) and avoid collisions (CA 0.74 vs 0.17 for a flocking-only policy).
As the authors review it, most agent-based collective-motion models give each agent exact neighbour positions and velocities, the vision-based occlusion models they discuss ignore agent-specific features such as shape, and learning-based approaches either addressed a single objective or assumed full information. Here an individual-level learning scheme yields two swarm-level behaviours at once (velocity alignment and collision avoidance) from the visual cues of elongated rectangular agents.
Key results
ShownWith alignment and collision avoidance weighted equally, the learned policy flocks nearly as well as a flocking-only policy and has far fewer collisions. A random walk does well on neither objective.
Over n = 30 simulations: CM (alignment) score 0.75 for the flocking-only model (ζ = 1) and 0.7 for the two-objective model (ζ = 0.5), vs 0.4 for random walk. CA score 0.74 for the two-objective model vs 0.29 for random walk and 0.17 for the flocking-only model. Scores are complements of the objectives, so a higher CA means fewer collisions. · 4.3 Baseline
ShownAgent-level training fitness and the swarm-level CMCA score both rise roughly logarithmically with the number of training samples and are tightly correlated. The authors read this as the individual objective carrying over to group behaviour.
Across n = 30 training runs, least-squares fits gave fitness = 0.189ln(samples) + 0.053 (R^2 = 0.989) and cmca = 0.196ln(samples) − 0.268 (R^2 = 0.972). The validation-stage Pearson correlation between the two is reported as R^2 = 0.971 (p-value 0.0057). Training converged after five validations with κ = 0.05, Ztraining = 10000000 and Zvalidate = 100000. · 4.2 CEL algorithm training; Figure legends
ShownThe trained policy's performance depends on swarm size and agent shape. The combined CMCA score peaks at an intermediate population size and falls as agents become more elongated; square agents score best.
The policy was trained at default settings (height:width 7/1) and evaluated on other settings, with n = 30 simulations each. Best flock size was P = 40. Higher height-to-width ratios lowered CMCA, and square agents scored highest. The authors attribute the shape effect to orientation-dependent noise in the subtended angle ψ. The text gives no values for the sensing-radius sweep. · 4.1 Setup; 4.4 Parameter sensitively analysis
ShownA random-forest surrogate cannot fully reproduce the learned policy: agreement stops improving even with hundreds of leaves. The authors take this to mean the policy cannot practically be explained as a set of rules.
Random forest models with 1000 trees were fitted to 1000000 random states queried from the policy. Accuracy plateaued at 93.37% beyond 320 leaf nodes or max depth 8, with 6.63% of cases treated as anomalies. A sharp gain between 70 and 80 leaf nodes coincided with telling 'keep heading' apart from 15-degree left or right turns. · 4.5 Explainablility of the model; 5 Discussion
SuggestedAlignment builds up gradually from a disordered start. In the two-objective model, collision avoidance declines slowly over the run, which the authors attribute to alignment drawing agents together. They compare the alignment dynamics to desert locust hopper-band formation.
Over n = 30 simulations, the CM index rises at a slowing rate. Collision rates for the CM and random-walk models settle quickly. The two-objective model's CA score falls moderately throughout the run and saturates at a value the authors describe as more than four times the CM model's. The hopper-band comparison is qualitative. · 4.3 Baseline
Methods, data and reproducibility
Rectangular agents in a 2D toroidal world sense neighbours within a metric radius only through visual cues: angular position, subtended angular area and their time derivatives. Collective Evolution Learning trains one shared, discretized Q-learning policy. A genetic algorithm finds good short action sequences for random local scenarios, and distance-weighted KNN decides where to sample next.
- Upper-view locust images with manually tagged bounding boxes, used to set the agent height-to-width ratio (mean 6.78, SD 1.48; rounded to 7/1) · Schistocerca cancellata and Schistocerca gregaria (locusts) · n = 50 images
- In silico locust-swarm simulations from the VCMCA simulator, used for training and for baseline and sensitivity evaluation · n = n = 30 training runs; n = 30 simulations per evaluated setting
- Randomly generated states queried against the trained policy, used to fit random forest surrogates (1000 trees) · n = 1000000 cases
- Methods & tools
- Collective Evolution Learning (CEL), VCMCA simulator, Q-learning, genetic algorithm, distance-weighted K-nearest neighbors, random forest
- Reproducibility gaps
- No software versions, hardware or run times are stated. The 50 locust images used for shape calibration are not linked. Default hyperparameters are in Table 1, whose values are not in the provided text.
Limitations (6)
- The model is two-dimensional; the authors note that extending it to 3D geometry would be more realistic. · stated by the authors
- All agents are identical and share one policy, while real swarm members differ in size, capability and decision-making. This heterogeneity could produce other behaviours, such as leader selection. · stated by the authors
- There is no direct comparison with Vicsek-type models. The authors argue it is infeasible because those models use different inputs, and adapting them to vision-based input would change their behaviour. · stated by the authors
- Simulated densities are an order of magnitude below recorded locust swarm densities. The authors restrict their focus to the early, emerging phase of flocking. · stated by the authors
- All evidence is in silico: the learned behaviour is never compared with recorded locust trajectories. The only baselines are a flocking-only ablation of CEL (ζ = 1) and a random walk. · noted in this summary
- The Discussion's robustness claim rests on a policy trained only at default settings. The text reports a peak at P = 40 and a shape trend but gives no values for the sensing-radius sweep, so that effect cannot be judged from the prose. · noted in this summary
Who should read:Read this if you model animal collective motion and want an agent-level learning scheme (Q-learning plus a genetic algorithm) that gets both flocking and collision avoidance from the vision-only cues of elongated agents.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- Vision-Based Collective Motion: A Locust-Inspired Reductionist Model
D. L. Krongauz, A. Ayali, G. A. Kaminka·bioRxiv·Jan 17, 2023
- A model of collective behavior based purely on vision
R. Bastien, P. Romanczuk·bioRxiv·Mar 28, 2019
- Visual processing and collective motion-related decision-making in desert locusts
I. Bleichman, P. Yadav, A. Ayali·bioRxiv·Sep 19, 2022
- Perception of Motion Salience Shapes the Emergence of Collective Motions
Yandong Xiao, Xiaokang Lei, … Xingguang Peng·Nature CommunicationsPeer-reviewed·Jun 5, 2024
- Visual social information use in collective foraging
David Mezey, Dominik Deffner, … Paweł Romańczuk·PLoS Computational BiologyPeer-reviewed·May 3, 2024
- Collective motion as a distinct behavioral state of the individual
D. Knebel, C. Sha-ked, … A. Ayali·bioRxiv·Jun 15, 2020