Automated discovery and curiosity-driven exploration of behavioral repertoires

Calvo and Baluška (2015) argued that the behavioral repertoire of eukaryotes, and the organization that sustains it, cannot be read in purely reactive terms¹. Calvo, Baluška and Sims (2016) described the opposing view, in which plant behavior reduces to fixed tropistic or nastic reflexes. They noted that behavioral flexibility is a marker of cognitive sophistication, and that hard-wired behavior lacks it². Levin (2023) added that repertoires can change radically within one individual, not only over evolutionary time³. On this reading, knowing what a system can do is itself a scientific question, and the extent of a repertoire is contested.

Budnikova and colleagues (2014) laid the foundation for automated model discovery that draws on shape-based data from a repository of documented experiments⁴. Lobo and Levin (2015) modeled gene, protein and metabolite regulatory networks with nonlinear partial differential equations, as the setting for an automated discovery system⁵. In both cases the search was aimed at explaining experimental morphology.

Etcheverry, Moulin-Frier, Oudeyer and Levin (2025) defined goal states as attractor states of a gene regulatory network. They reported that traditional screening can be very inefficient at finding the range of such states. They proposed intrinsically-motivated goal exploration processes (IMGEP) instead⁶. They compared random search with curiosity-driven exploration using threshold coverage, which measures the area of the space reached at a given precision⁷.

Blackiston and colleagues (2023) described xenobots as autonomous, adaptive and showing a diverse array of behavioral repertoires, and left open whether they are intelligent⁸. Gumuskaya and colleagues (2025) said the rules governing Anthrobot motion were unknown. They proposed experiments in enriched environments to probe their repertoire⁹.