Neural Circuit Architectural Priors for Rat Locomotion
TL;DR A controller wired like mammalian spinal locomotion circuits (Rat NCAP) drives a 38-DoF simulated rat to stand, walk, trot and bound in one model, switching gaits as its speed command rises, without imitation learning. It reaches about twice the return of the earlier Quadruped NCAP on the same body.
What they do
Simulated animals that move naturally usually imitate recorded kinematics, and generic networks are hard to relate to real neural circuits. The authors build Rat v2, a 38-DoF MuJoCo rat in a GPU-accelerated JAX framework, and Rat NCAP, a controller whose structure follows mammalian spinal locomotion circuits. They train it with simple velocity-tracking rewards and compare its gaits and neural activity with rat data.
How it works
- Body. Rat v2 refines the Merel et al. (2019) rat: torque actuators, recalibrated joint ranges and force limits, and paw sensors that report contact force.
- Circuit. A fixed, hand-tuned rhythm generator of flexor/extensor half-centers linked by commissural interneurons; a brainstem command that maps target speed to rhythm frequency and amplitude; learned afferent-feedback and pattern-formation modules (delays, synergies, sign-constrained motor weights). It controls the 31 limb and trunk DOFs.
- Interfaces. Circuit output becomes torque through torque, position, position-stiffness or flexor-extensor (antagonistic spring) control.
- Training. Augmented Random Search on GPU at six fixed target speeds from 0 to 1 m/s, about 36 h per run on one L40S.
Results
- One model stands, walks, trots and bounds, and goes walk to trot to bound when the command rises mid-episode, though trained only at fixed speeds.
- As in rats, stance shortens with speed while swing stays roughly constant, and step frequency lies in the reported range; step length is too long at intermediate speeds.
- About twice the return of Quadruped NCAP on Rat v2. PPO- and ARS-trained MLPs reach comparable return but keep stepping at 0 m/s with irregular footfalls.
- Flexor-extensor control gives the highest return and best speed tracking, torque control the lowest. Scaling V0v and V0d activity to zero replaces trot with bound, matching mouse knockouts; trot holds until k ≈ 0.35 and gives way to bound by k ≈ 0.1.
Limitations
Simulation only, trained on flat ground. Step length at intermediate speeds and knee and elbow flexion during swing do not match rats. Rhythm-generator weights are not identifiable from data, and the learned modules find plausible rather than unique solutions.
Videos
Figures




Abstract
Animals generate behavior through continuous interactions between their neural circuits, bodies, and environment. Embodied simulations study these interactions by testing neural circuit models in biologically realistic settings. Recent efforts have simulated 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, an efficient and flexible biomechanical body 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 model learns to locomote at different speeds and transition between gaits without imitation learning, producing locomotion statistics, kinematics, and cell-type-specific neural activity consistent with experimental data. 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.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
