NVIDIA Research Announces Three Neural Technologies to Transform Robot Learning

NVIDIA Research announced three neural technologies to transform robot learning on September 25, 2025. NeRD achieved a cumulative reward error of less than 0.1% in a 1,000-step policy evaluation on the ANYmal quadruped robot. Dexplore improved the success rate by about 20% on the Inspire hand. VT-Refine’s tactile simulation is based on the GPU-based TacSL library integrated with Isaac Lab.

For robot developers, the key point is that the methods are presented with clear usage and validation conditions for three challenges that are difficult to address in real-world environments. According to NVIDIA Research, NeRD predicts future states of robots in contact-rich settings in a form that can be incorporated into existing simulators; Dexplore learns dexterous manipulation from monocular depth images and a small number of user-provided goals; and VT-Refine is designed to use vision and touch for bimanual assembly, with additional training in simulation after pretraining on about 30 demonstrations.