Intrinsic Goals for Autonomous Agents: Hierarchical Exploration Captures Future Volitional Action in Drosophila

Project · arXiv · Code

Can intrinsic motivation explain spontaneous animal behavior? We train hierarchical agents in a biomechanically realistic fly body using 20 intrinsic motivation objectives spanning curiosity, entropy, and competency methods. The agents select among seven motor primitives and learn without external rewards or training on fly data. We then compare their spontaneous behavior and internal activity with fly recordings in a matched tethered-ball assay.

Among the objectives tested, DIAYN provides the strongest joint match to individual differences in fly turning behavior, internal activity predictive of future turns, and the sum-difference relationship between bilateral activity and locomotion. Its learned skills capture the diversity of exploratory strategies across flies, suggesting that individual behavioral differences may support state-space coverage at the population level.

Recommended citation: Declan Kutscher, Reece Keller, Alyn Kirsch, Felix Pei, Aran Nayebi. (2026). "Intrinsic Goals for Autonomous Agents: Hierarchical Exploration Captures Future Volitional Action in Drosophila." arXiv preprint arXiv:2610.34638.