A latent control model for realistic rodent motion

  1. Aidan Sirbu1,2,3,✉
  2. Charles Y. Zhang4
  3. Yuanjia Yang6,7
  4. Bence P. Ölveczky5
  5. Talmo D. Pereira6
  6. Blake A. Richards1,2,3,8,9,10
  1. 1Google, Paradigms of Intelligence Team
  2. 2Mila – Quebec AI Institute
  3. 3School of Computer Science, McGill University
  4. 4Harvard University
  5. 5Department of Organismal and Evolutionary Biology, Harvard University
  6. 6Salk Institute for Biological Studies
  7. 7University of California San Diego
  8. 8Department of Neurology and Neurosurgery, McGill University
  9. 9Montreal Neurological Institute, McGill University
  10. 10Learning in Machines and Brains Program, CIFAR

✉aidan.sirbu@mail.mcgill.ca

NeurIPS 2026

Abstract

The development of embodied models of animal behavior is a major goal of the neuro-AI community. Embodied models will enhance our ability to make direct comparisons between biological and artificial agents. To make such comparisons effective, though, we need embodied models that actually move like animals. Previous work has developed virtual animal bodies and techniques for training realistic motion through imitation. However, to-date, there is no system for embodied animal modeling that produces natural movements both during trained, goal-directed behavior, and during exploration driven by random generation. Here, we introduce Simulated Control via Augmented Motor Priors for Embodied Reinforcement learning (SCAMPER), a system for training an embodied virtual rat model, enabling natural movements during both goal-directed behavior and random exploration. SCAMPER uses a variational architecture with two distinct priors, a motion prior for control and a behavior prior for exploration. We show that combining both of these priors allows SCAMPER to generate realistic behavior during exploration and exploit the task structure effectively. As well, SCAMPER can learn from sparse rewards as a result of the diverse movements induced by the behavior prior, allowing the system to model realistic sparse reward tasks from experimental neuroscience. Altogether, SCAMPER provides the neuro-AI community with an effective new system for training embodied models of rodents on realistic tasks.

Overview

SCAMPER overview figure
Training stages of SCAMPER. (1) The expert imitator is trained with mocap data. (2) The latent controller is trained via online distillation. (3) The behavior prior is distilled from the frozen encoder. Finally, (4) the model can be trained with RL to perform different tasks.

Imitation & Random Motion Generation

Imitation. Rolling out the trained expert imitator on the imitation task. Colored rodents are the controlled agents, the translucent gray overlay on each rodent is the mocap target.
Behavior prior sampling. Naturalistic and behaviorally diverse motion sequences are generated by stochastically sampling the behavior prior with low noise.
Motion prior sampling. Stochastically sampling the motion prior with low noise results in locally-plausible, but behaviorally narrow motion that does not reproduce the behavioral diversity of the training data. Sampling with high noise is able to match the diversity of the training data with the cost of significant jitter, decreasing the naturalism of the motion.

Locomotion Task

Walk. SCAMPER, MIMIC-MJX, and the from-scratch model being tested on the locomotion task.
Trot. SCAMPER, MIMIC-MJX, and the from-scratch model being tested on the locomotion task.
Gallop. SCAMPER, MIMIC-MJX, and the from-scratch model being tested on the locomotion task.

Rearing Task

Rearing. SCAMPER, MIMIC-MJX, and the from-scratch model being tested on the rearing task. Task success corresponds to the agent being able to maintain its head within the red target region for a full second. Only SCAMPER is able to solve the task.

Arena Forage Task

In order to effectively solve the arena foraging task, SCAMPER dynamically gates between stochastic sampling of its behavior prior (exploration) and residual control of its motion prior (exploitation). The task policy network outputs a gate logit, $\beta_t$, which is binarized to $b_t \in \{0,1\}$ via hysteresis thresholding. This mechanism allows SCAMPER to generate realistic motion during early exploration in absence of a reward signal. As the agent discovers treats and begins receiving a reward signal, SCAMPER begins relying on residual control of the motion prior to exploit the task structure.

Exploration. During early training, SCAMPER relies heavily on its behavior prior to drive random exploration.
Early Exploitation. As SCAMPER begins to collect treats and sparsely obtain reward, it rapidly switches between residual control of its motion prior and stochastic sampling of its behavior prior in order to begin exploiting the task structure.
Exploitation. During late training, SCAMPER mostly relies on residual control of its motion prior to exploit the task structure effectively.
Baselines. Fully trained baseline models (MIMIC-MJX, from-scratch model) on the arena foraging task. MIMIC-MJX results in sluggish and jittery motion. The from-scratch model exhibits highly unnaturalistic behavior to solve the taks.

BibTeX

@inproceedings{sirbu2026scamper,
  title     = {{SCAMPER}: A Latent Control Model for Realistic Rodent Motion},
  author    = {Sirbu, Aidan and Zhang, Charles Y. and Yang, Yuanjia and {\"O}lveczky, Bence P. and Pereira, Talmo D. and Richards, Blake A.},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}