We introduce SCAMPER (Simulated Control via Augmented Motor Priors for Embodied Reinforcement learning), a framework for training a virtual rat that moves naturally both during goal‑directed tasks and random exploration.
Previous work has built virtual animal bodies and used imitation learning to produce realistic motion, yet no system has achieved natural movement across both directed behavior and stochastic exploration.
SCAMPER adopts a variational architecture with two distinct priors: a motion prior that supplies controllable motor trajectories, and a behavior prior that generates diverse exploratory actions. This dual‑prior design enables the agent to balance task performance with behavioral richness.
Results show that the combined priors yield realistic exploratory trajectories and allow the agent to exploit task structure effectively. Moreover, the diversity induced by the behavior prior lets SCAMPER learn from sparse rewards, mirroring the conditions of many neuroscience experiments.
Overall, SCAMPER offers the neuro‑AI community a powerful new tool for training embodied rodent models on realistic tasks, facilitating direct comparisons between biological and artificial agents.
Review