Animals generate behavior through continuous interactions between their neural circuits, bodies, and environment. Embodied simulations provide a means to study these interactions by testing neural circuit models in biologically realistic settings. Recent efforts have developed embodied simulations of diverse animals, and emerging work on biologically grounded neural architectures has begun to integrate circuit connectivity and single-unit dynamics from experimental data into artificial neural networks, introducing stronger inductive biases and closer mechanistic alignment with biology. How can we develop embodied simulations that enable the study of increasingly complex bodies and neural circuits? In this work, we address the challenge through coordinated advances in the simulated body, neural architecture, and training infrastructure. We introduce Rat v2, a biomechanical model that provides an efficient and flexible tool for simulating rat behavior. To control this body, we develop Rat NCAP, a biologically grounded neural architecture for locomotion based on mammalian spinal circuits that incorporates cell-type-specific connectivity and a novel flexor-extensor neuromechanical interface. To optimize circuit parameters, we develop GPU-accelerated simulation and training infrastructure. The integrated system learns to locomote at different speeds and transition between gaits without imitation learning, producing loco motion statistics, kinematics, and cell-type-specific neural activity consistent with experimental data. Systematic analyses reveal that the neuromechanical interface substantially affects trainability and gait quality, and in silico cell-type ablations reproduce known gait phenotypes and predict gait changes under graded perturbations. Together, these findings support the view that embodied simulations at an intermediate level of abstraction can simultaneously reproduce complex behavior and yield mechanistic insights into the neural circuits that generate it.
Rat NCAP learns to walk, trot, and bound within a single model, producing naturalistic stepping and speed-dependent gaits. Prior work in this rat relied on imitation learning to obtain naturalistic behavior [Aldarondo et al., 2024, Zhang et al., 2025] or produced visibly unrealistic movement without it [Merel et al., 2019]. Rat NCAP achieves comparable naturalism using only simple velocity rewards, demonstrating the powerful inductive bias of biological neural circuits.
Increasing brainstem command within a single episode reveals distinct gait patterns at each speed, consistent with speed-dependent recruitment of distinct locomotor patterns. Notably, gait transitions are not trained and reflect zero-shot generalization.
The neuromechanical interface determines not only trainability, but also the qualitative character of learned gaits. Despite similar asymptotic reward, the flexor-extensor interface produces qualitatively different locomotion: it uses substantially more negative mechanical power than the other interfaces, storing and returning energy through its antagonistic spring structure analogous to tendon compliance. This compliance is visible in joint trajectories: position and position-stiffness produce sharp arcs with brief ground contact, while flexor-extensor produces rounded trajectories with extended stance as the paw deforms along the ground.
A central advantage of a biologically grounded circuit model is the ability to perform targeted in silico ablations of identified cell types, which is an analysis not possible with black-box controllers. In mice, ablating V0v neurons eliminates trot, and additionally ablating V0d neurons leaves bound as the only coordinated gait [Bellardita and Kiehn, 2015]. Rat NCAP reproduces this combined phenotype: scaling V0v and V0d activity to zero replaces trot with bound. In contrast, scaling V3f and V3a activity to zero leaves trot intact, so not all commissural populations are equally critical for trot. Together, these results show that RG cell-type phenotypes persist in an embodied setting and yield experimentally testable predictions.