TorqueNet, by NVIDIA

A reinforcement learning-based neural actuator that bridges the joint-level simulation-to-reality gap in motion execution. Trained on IsaacLab to learn optimal joint torques from position commands, outperforming traditional PD controllers.