Mathis Hocke, Andreas Gerken, David Bierbach, Jens Krause, Tim Landgraf
4 min
This paper introduces a groundbreaking framework that uses biomimetic robotic fish (RoboFish) to actively test and validate computational models of fish shoaling behavior. Traditional methods evaluate models by passively comparing simulated trajectories to real data—static, offline, and blind to dynamic interactions. Here, researchers train reinforcement learning (RL) policies in simulation using candidate fish behavior models, then deploy them on real RoboFish to interact closed-loop with live guppies. By measuring how real fish respond compared to simulated predictions, they quantify the sim-to-real gap and rank model fidelity. The neural network-based model outperforms rule-based ones, proving the approach can distinguish good from bad hypotheses.
Biomimetic robots like RoboFish mimic live fish in appearance and motion, allowing seamless integration into real shoals. By replacing one fish with a controllable robot, researchers probe social interactions: perception (do fish notice the robot?), response (do they react naturally?), and coordination (does group behavior persist?). This goes beyond open-loop playback (pre-recorded motions), where animals can't influence the robot. Closed-loop interaction is crucial—fish evasion immediately alters robot path, revealing adaptive dynamics invisible in static stats.
The paper contrasts this with prior 'biomimicry gap' work, which tests if models transfer to robots without embodiment artifacts. Instead, they use robots as standardized probes: all models face identical real-world conditions, isolating behavioral inaccuracies.
The core innovation: RL policies operationalize model hypotheses. Policies are trained to guide a simulated 'focal fish' to goal locations (e.g., join shoal, avoid walls) while interacting with virtual fish governed by candidate models:
Zero-shot transfer to RoboFish enables real fish interactions. Success metrics (goal-reaching time, inter-fish distances, alignment) form behavioral distributions. The sim-to-real gap uses Wasserstein distance (optimal transport metric) between simulated vs. real distributions—small gap means the model accurately predicts live responses.
Wasserstein distance captures distributional shifts holistically, unlike mean errors. CNN model shows smallest gaps across metrics, validating its superior fidelity. Rule-based models fail on nuances like wall interactions or precise alignment, exposed only through embodied testing.
This mirrors the scientific method: hypothesize rules → derive predictions (via RL) → falsify empirically. Interactive probes scale to complex collectives, where offline stats can't probe causality.
Evaluating animal behavior models is key for collective dynamics (shoaling, flocking), decision-making, and bio-inspired robotics/swarm AI. This framework uncovers deficiencies static benchmarks miss, provides quantitative rankings, and generalizes beyond fish (birds, ants). Future: multi-robot swarms, online model updates from real data.
Understanding and modeling animal behavior is essential for studying collective motion, decision-making, and bio-inspired robotics. Yet, evaluating the accuracy of behavioral models still often relies on offline comparisons to static trajectory statistics. Here we introduce a reinforcement-learning-based framework that uses a biomimetic robotic fish (RoboFish) to evaluate computational models of live fish behavior through closed-loop interaction. We trained policies in simulation using four distinct fish models-a simple constant-follow baseline, two rule-based models, and a biologically grounded convolutional neural network model-and transferred these policies to the real RoboFish setup, where they interacted with live fish. Policies were trained to guide a simulated fish to goal locations, enabling us to quantify how the response of real fish differs from the simulated fish's response. We evaluate the fish models by quantifying the sim-to-real gaps, defined as the Wasserstein distance between simulated and real distributions of behavioral metrics such as goal-reaching performance, inter-individual distances, wall interactions, and alignment. The neural network-based fish model exhibited the smallest gap across goal-reaching performance and most other metrics, indicating higher behavioral fidelity than conventional rule-based models under this benchmark. More importantly, this separation shows that the proposed evaluation can quantitatively distinguish candidate models under matched closed-loop conditions. Our work demonstrates how learning-based robotic experiments can uncover deficiencies in behavioral models and provides a general framework for evaluating animal behavior models through embodied interaction.