Neural Robot Dynamics (neural-robot-dynamics.github.io)

Researchers introduced NeRD (Neural Robot Dynamics), a learned, robot-specific dynamics backend that replaces only the application-agnostic low-level forward dynamics and contact solvers inside a conventional robotics simulator. Targeting articulated rigid-body robots with contacts, NeRD uses a robot-centric, spatially invariant state representation and a hybrid prediction framework: classical high-level simulator logic remains, while neural networks predict the low-level dynamics and contact responses from a short history window. The integrated NeRD solver runs stably and accurately for thousands of steps, and a single NeRD model per robot instance generalizes across diverse contact configurations and spatial locations. NeRD’s key implications for AI/ML and robotics are practical and technical: it enables task- and environment-general neural simulators (so you don’t need per-task retraining), supports policy learning entirely inside a neural engine with rollouts that closely match analytical simulators, and can be fine-tuned on real-world data to reduce sim-to-real gaps. Empirically, NeRD models generalized on benchmarks from double pendulums to legged and manipulator tasks (ANYmal, Franka, Ant, Cartpole), enabled zero-shot real deployment of a Franka reach policy, and showed high fidelity to ground-truth simulators—demonstrating a promising route to faster, more adaptable simulation and more reliable sim-to-real transfer.