PaperlayerView exact revision on arXiv

Source-linked AI summary

The value of information and the emergence of communication in evolved e-puck robots

Fernando Montes-Gonzalez

arXiv:2610.13118v1cs.ROcs.MAcs.NE

TL;DR

This paper asks when communication emerges between autonomous agents and what role an innate action-selection architecture plays. It studies two e-puck robots foraging in a validated model, then finds that communication disappears when its information is available through another channel and that harder tasks do not make communication more common.

  • Problem

    The paper asks under what conditions communication emerges between autonomous agents, a question that is difficult to study in animals because the origin of signaling can rarely be observed.

    Source: §1 · pp. 1–2

  • Method

    The authors build an open, validated two-e-puck system with a one-bit signal and compare evolved control in a foraging task.

    Source: §1 · p. 3, §6 · p. 19

  • Results

    Communication disappears when its information is available through another channel, appears in most but not all runs otherwise, and becomes no more frequent in harder tasks.

    Source: §6 · p. 19

  • Takeaways & Limitations

    The findings link communication emergence to whether its information is independently available, while greater task difficulty does not make communication more common.

    Source: §6 · p. 19

    Each main condition has three runs, enough to show large consistent differences but not to estimate the frequency of each communication strategy accurately.

    Source: §5.4 · p. 19

Abstract

from arXiv · show

We study when communication emerges between autonomous agents and what role an innate action selection architecture plays. Two e-puck robots, reconstructed in PyBullet from their Webots model and validated against it, must find a feeding zone that can only be detected when a robot passes over it. Each robot can emit a one-bit signal whose direction is detected by the receiver. Control is provided by a GRU recurrent network with a residual path, either operating alone or as a channel within an Embedded Action Selection Architecture (EASA). Fitness measures only the time the robots spend over the food. When the robots receive a compass pointing to the food, use of the signal disappears completely. Without a compass, in a small arena, three independent runs achieved similar performance using different strategies: a referential signal, a permanent beacon, or no use of the signal. As the arena expands, the strategy based on a referential signal generalizes best. In contrast, evolving directly in the large arena does not increase the frequency of communication, but it does separate the strategies, and the referential signal clearly outperforms solitary searching. Within EASA, innate reflexes constrain the network and reduce performance. As its priorities evolve, evolution consistently weakens those reflexes. The results suggest that communication emerges in proportion to the value of the information it transmits, and that an innate scaffolding designed for isolated agents can be an impediment in a social task.

1. Introduction

The introduction asks when communication emerges, emphasizing that signals are useful when they provide information the receiver cannot obtain independently. It presents a validated two-robot foraging system to test information availability and controller architecture, and reports that signal use depends on information value while innate reflexes can constrain social behavior.

1. Introduction

  • Two e-puck robots search for food detectable only by passing over it, using a direction-detected one-bit signal and GRU control alone or within EASA.The robots receive no food-position information, and fitness rewards only time over food.Source: §1 · p. 3
  • The study varies information need through compass availability and arena size, and compares network-only control with basal-ganglia architectures whose priorities may be fixed or evolved.Each condition is independently replicated and evaluated with paired ablations, behavioral diagnostics, and robustness tests.Source: §1 · p. 3
  • The work introduces an open two-e-puck model with one-bit communication, reconstructed from Webots in PyBullet to address simulator-validation concerns.The model is designed for replication and extension by other groups.Source: §1 · pp. 2–3
  • When information provided by a signal becomes available through another channel, signal use disappears; otherwise, communication appears in most, but not all, runs.The study rewards only time spent over food, not signal transmission or approach.Source: §1 · p. 3
  • A more difficult task does not increase communication frequency, but favors the referential signal over other strategies.The introduction also reports that independent runs can reach similar performance through different solutions.Source: §1 · p. 3
  • In the social task, innate basal-ganglia reflexes constrain the learned network, and evolving priorities tends to weaken those reflexes.Source: §1 · p. 3

2. Related work

Prior work examines how communication can emerge with or without predefined signaling channels, and how action-selection architectures and simulation choices shape studies of robot behavior.

2.1 Evolution of communication in robots

  • Coordination without a dedicated channel can become communicative, while evolved navigation robots developed non-trivial signals that co-adapted with behavior.Source: §2.1 · p. 4
  • Relatedness and conflicting interests were studied as conditions affecting signaling, while non-interaction can remain evolutionarily stable even when communication could benefit both parties.The study adopts homogeneous teams and team-level selection to isolate informational value.Source: §2.1 · p. 4

2.2 Action selection and basal ganglia in robots

  • Less work has examined action-selection architectures when a channel is an evolution-acquired network and the task requires coordination with another agent.EASA is described as an open reimplementation of a model for robot behavior.Source: §2.2 · p. 5

2.3 Simulation and design of the fitness function

  • Simulation can expose evolved controllers to artifacts absent from the real world, motivating validated robot models such as the e-puck reconstruction in PyBullet from Webots.The reconstruction's components were verified before use.Source: §2.3 · p. 5
  • Implicit fitness measures task outcomes, whereas explicit functions reward selected behaviors and can impose the designer's solution.For communication studies, bonuses for signaling or approaching a partner make signaling part of the objective; this work measures time over food.Source: §2.3 · pp. 5–6

3. Methods

The methods describe a Webots-derived PyBullet e-puck task for testing recurrent controllers with a one-bit directional signal, either alone or within EASA, in shared-genome teams. They also specify the fitness, champion selection, and evaluation procedures, including held-out episodes, signal manipulations, behavioral measures, and robustness tests.

3.1 Robot model and environment

  • The model was validated against Webots' kinematics; straight-line speed was 0.1255 m/s versus expected 0.1256 m/s, and simulation steps occurred every 66.7 ms.The arena was a 1 m square, the food zone a 10 cm-radius circle, and episodes lasted 60 s.Source: §3.1 · pp. 6–7

3.2 Signal

  • The signal is a content-free bit whose reception provides direction and distance-decaying strength, but no food location; only evolution can associate it with food.Packets arrive one control step after transmission and have unlimited range.Source: §3.2 · p. 8

3.3.1 Recurrent network with residual path (Res-GRU)

  • A 16-unit GRU receives 17 inputs and uses a residual input-output path for memory-independent reactions, producing wheel commands, signal transmission, and an EASA control bid.Inputs include proximity and floor readings, signal perception, and fear and hunger motivations.Source: §3.3.1 · p. 8

3.3.2 EASA Architecture

  • EASA selects among innate exploration and avoidance channels and the Res-GRU using channel salience, with thalamocortical persistence and motor gating.Communication is not innate: only the network can transmit or receive the signal.Source: §3.3.2 · p. 9
  • The study uses three EASA variants: manually weighted safety channels, an added innate feeding channel, and a variant that evolves 17 selection parameters with the network.The evolved parameters comprise 12 salience weights, three persistence weights, and D1 and D2 dopamine levels.Source: §3.3.2 · pp. 9–10

3.4 Evolution and fitness function

  • Teams share a genome, and evolution uses populations of 48 for 150 generations, with 10 elites, crossover from the top 25, and six evaluation episodes per individual each generation.Offspring receive Gaussian noise with probability 0.2 and standard deviation 0.12 in all genes.Source: §3.4 · p. 10
  • Fitness was the fraction of each episode that robots spent over food, averaged across both robots and six episodes, with episodes lasting 60 seconds and scores ranging from 0 to 1.Fitness did not reward emitting signals or approaching the food or partner.Source: §3.4 · pp. 10–11
  • The compass condition added food direction and proximity inputs, while the large-arena condition used a 1.5 m square and 90-second episodes.The compass supplied information unavailable in the other conditions.Source: §3.4 · p. 11

3.5 Evaluation

  • Champions were chosen from the best of the last ten generations using 12 validation episodes, then tested in 100 new episodes with normal, disabled, and randomized-direction signals.The randomized version preserved whether a partner emitted a signal but removed its location information; the three versions used identical episodes for Wilcoxon signed-rank comparisons.Source: §3.5 · p. 11
  • Disabling signals requires caution because removing a near-constant input may disrupt the network independently of the signal’s information, so results are interpreted with randomized signals and signal specificity.Comparisons between conditions with three runs used Welch’s t-test on each run’s average.Source: §3.5 · p. 11
  • Across 30 test episodes, behavioral assessment measured food occupancy, joint occupancy, departures lasting at least half a second, signal specificity on and off food, EASA active channels, and avoidance when the partner was within 15 cm.Source: §3.5 · p. 12

4. Results

In the small arena, the GRU outperformed EASA, while different communication strategies achieved similar performance. A compass eliminated signaling; as arenas grew, referential signaling generalized better than beaconing or solitary search, while EASA’s innate influences were weakened during evolution.

4.1 Performance by condition

  • In small-arena tests, GRU champions averaged 0.587 fitness versus 0.362 for evolved EASA, with a significant difference between conditions.The comparison covers 100 test episodes; Welch’s t-test gave p = 0.0001.Source: §4.1 · p. 12, Table 3 · p. 13

4.2 Signal Function

  • Signal ablations showed that communication improved fitness in some runs but had no effect in others.Turning off the signal reduced fitness by 0.128 in one GRU run; randomizing direction reduced it further, while another GRU run was unaffected by all three signal conditions.Source: §4.2 · pp. 13–14, Figure 5 · p. 14

4.3 Communication strategies across runs

  • GRU runs achieved the same performance via a referential signal, individual search without a signal, or a beacon emitting almost always away from food; all three kept both robots on food similarly.With the referential signal, the robot emitted in 96% of control steps spent on food and 1.3% of those spent off it.Source: §4.3 · p. 14

4.4 Compass Effect

  • With a compass, signal transmission fell below 1.5% by generation 65 and reached zero by generation 90; test fitness was identical across signal conditions.The robots arrived together in 100% of episodes and spent 93% of their time on the food.Source: §4.4 · p. 15

4.5 Restrictions on innate action selection

  • In EASA, innate channels took control for about a third of steps, almost always away from food, disrupting the network’s search despite selection persistence.The avoidance reflex also treated a nearby partner as an obstacle, though only 5 of 46 food departures were preceded by avoidance.Source: §4.5 · pp. 15–16

4.6 Evolved priorities of the basal ganglia

  • Across all three evolved EASA runs, evolution reduced innate influence, including the obstacle-detection weight on avoidance from 0.9 to 0.53–0.79.The weights of obstacle inhibition of the network and hunger drive on exploration also decreased in the reported ranges.Source: §4.6 · p. 16

4.7 Robustness and large arena

  • On transfer to the large arena, the referential-signal run lost 5% fitness, versus 23% for the beacon and 36% for noncommunication; EASA variants lost 23%–45%.Motor and sensor noise had effects below 15% across runs, while the compass network showed no fitness reduction.Source: §4.7 · p. 16, Figure 7 · p. 17
  • In the large arena, referential signaling outperformed solitary search by 0.165 fitness, while beaconing scored 0.229 and fell below noncommunication.For champions evolved in the small arena, the referential-signal advantage was about 15 seconds over the food per robot (p = 0.0003).Source: §4.7 · p. 17, Table 5 · p. 17

5. Discussion

The discussion finds that communication depends on information the receiver cannot otherwise obtain, while task difficulty distinguishes useful from unhelpful signals. Innate reflexes can constrain social behavior, and the authors note limits in runs, simulations, and experimental scope.

5.1 Communication follows the value of information

  • With a compass, communication disappeared; without one, it appeared in most but not all runs, taking different forms.The communicating runs included referential signals and beacons.Source: §5.1 · pp. 17–18
  • Communication depends on information unavailable to the receiver; privileged target information can make it redundant and complicate claims that it evolved.The authors note that bonuses or rules can enforce communication, making it difficult to tell whether it emerged through evolution.Source: §5.1 · pp. 17–18

5.2 When the task is more difficult, what the signal says matters

  • A larger arena did not increase communication frequency, while the small-arena strategies showed similar performance.In the small arena, referential signaling, beaconing, and solitary search performed similarly.Source: §5.2 · p. 18
  • In the large arena, referential signaling outperformed solitary search, while beaconing performed worse than no communication; robustness tests preserved this order.The referential signal retained value as search became more difficult, whereas a partner-location beacon could lead the receiver astray.Source: §5.2 · p. 18

5.3 Innate scaffolding in a social task

  • EASA variants performed worse and were less robust than GRU; avoidance and innate channels interfered with seeking and approaching a partner.The avoidance reflex treats the partner as an obstacle when the robots need to be together.Source: §5.3 · pp. 18–19
  • When priorities could evolve, evolution weakened avoidance, protected the network from obstacle inhibition, and reduced exploration, shifting control toward learned behavior.Source: §5.3 · p. 19

5.4 Limitations

  • Three runs per main condition show consistent differences but cannot accurately estimate strategy frequency; behavioral conclusions use champions, and results are from simulation.The simulation model was validated against Webots, and runs used one machine because simulation physics differ across computers.Source: §5.4 · p. 19
  • The study uses one task, network form, and evolutionary algorithm, and tests only one way of combining reflexes with a learned network.Source: §5.4 · p. 19

6. Conclusions

The authors present an open, validated two-e-puck system for studying one-bit communication. Their results show that communication disappears when its information is available elsewhere, while harder search distinguishes useful signals and innate reflexes constrain learned control.

  • The open, validated two-e-puck system supports study of the emergence of a one-bit signal.Source: §6 · p. 19
  • Communication disappeared when its information was available through another channel, but appeared in most runs when it was not.Source: §6 · p. 19
  • A harder task did not make communication more frequent, but referential signals retained value relative to less informative strategies.Source: §6 · p. 19
  • In a social task, innate basal-ganglia reflexes constrained the learned network, and evolution consistently shifted control toward that network.Source: §6 · p. 19
  • A next step is to test basal-ganglia selection among learned proposals and verify the results with physical robots.Source: §6 · p. 19

Declarations

The declarations report no funding or conflicts of interest and state that the study involved no human participants or animals. Source code, scripts, and reproducibility datasets are publicly available.

Funding

Conflict of interest

Ethical approval

Data and code availability

Loading 2610.13118v1…