NVlabs/neural-robot-dynamics

What it solves

NeRD addresses the challenge of predicting future states for articulated rigid bodies (robots) under contact constraints. It aims to replace traditional, computationally expensive low-level dynamics and contact solvers in analytical simulators with a learned approach that maintains accuracy over long horizons.

How it works

NeRD implements learned, robot-specific dynamics models that function as interchangeable backend solvers. It utilizes a robot-centric and spatially-invariant simulation state representation to predict motion. The system is integrated into NVIDIA Warp, allowing users to switch between traditional analytical dynamics (like Featherstone or XPBD) and neural dynamics via a unified integrator interface.

Who it’s for

This project is designed for robotics researchers and engineers working on physics simulation, robot dynamics modeling, and reinforcement learning (RL) for articulated robots.

Highlights

• Interchangeable Backend: Can be plugged into NVIDIA Warp as a replacement for standard physics solvers.
• Neural Dynamics: Replaces low-level contact solvers with learned models to predict future states.
• RL Integration: Supports training RL policies directly within a simulator powered by NeRD dynamics.
• Pretrained Models: Provides ready-to-use models for environments like Cartpole, Ant, and ANYmal.