Simulation to reality
FlashNeRD Divides Contact-Rich Simulation Between Neural and Analytical Models
A hybrid architecture addresses NeRD’s fixed-contact and object-coupling limits, with its authors reporting gains in dynamics rollouts, policy training and inference-time planning.
FlashNeRD keeps collision detection analytical, uses a learned model for the robot’s numerical dynamics, and couples that model in both directions to objects handled by analytical solvers. Its authors describe an encoder that accepts contacts wherever they occur, removing NeRD’s restriction to predefined contact points, while a parallel streaming architecture is intended to make each prediction faster and more accurate.
In experiments spanning five robots, the authors report faster and more accurate dynamics. For three robots, FlashNeRD’s dynamics model was up to 55× faster than an optimized NeRD and remained more accurate over long rollouts. These figures compare simulation models and are claims from the FlashNeRD preprint rather than independent measurements.
The reported acceleration also carried into control workloads. PPO trained an ANYmal locomotion policy in 34 seconds, which the authors measured as 3.8× faster than the analytical simulator and 2.7× faster than an optimized NeRD. With DIAL-MPC, a sampling-based model-predictive control method, the robot climbed all six test platforms, compared with one for fixed-contact NeRD, while planning took approximately half as long as with the analytical simulator. Cube-reorientation policies trained with FlashNeRD completed within 2% of the target count achieved by simulator-trained policies.
The available evidence is the authors’ preprint abstract and describes simulation experiments; it does not establish independent confirmation or physical-robot validation.