NVIDIA Introduces ADEPT: Zero-Shot Real-World Transfer of Sim-Trained Visuo-Tactile Policies
dr_alphalyrae · x · 2026-09-01
NVIDIA, in collaboration with SharpaRobotics, introduces ADEPT, a pre-training and post-training paradigm that uses RL entirely in simulation to train visuo-tactile policies, followed by zero-shot deployment in the real world. The research utilizes IsaacLab to simulate five tactile sensors at the fingertips and successfully transfers policies to two different arm-hand combinations without tuning any RL hyperparameters or reward functions.
More from Embodied
- Using an iPhone 13 mini as a Reinforcement Learning Environment — willcb · 2026-09-01
- Avatar with Whole-Body Touch, Spatial Hearing, and Physical Voice in One Simulation — bhartsb · 2026-09-01
- Unitree G1 performs aerial flips in forest powered by BeyondMimic — CyberRobooo · 2026-09-01
- SpaceX Removes 12.1 Tons of Debris from Boca Chica Beach — DimaZeniuk · 2026-09-01
- Chinese Humanoid Robots Dominate IFA: UBTECH, Unitree, and More — CyberRobooo · 2026-09-01
- Microduck Training: From Collision Avoidance to Fortnite Dances — tristanbob · 2026-09-01