TacSL: A Library for Visuotactile Sensor Simulation and Learning

NVIDIA's open-source GPU library for simulating vision-based tactile sensors and training policies that use them.

TacSL was proposed by NVIDIA's Iretiayo Akinola, Yashraj Narang, and colleagues, posted to arXiv in August 2024, and later published in IEEE Transactions on Robotics. It uses the GPU inside NVIDIA's Isaac simulators to simultaneously generate visuotactile images (the deformation image an internal camera would capture on a sensor's gel pad, as with GelSight-style sensors) and contact force distributions, which the paper reports as more than 200 times faster than the best prior method — necessary for training tactile-input policies at large parallel scale. The library also includes contact-rich training environments, such as peg-in-hole insertion, and an asymmetric actor-critic distillation (AACD) algorithm for learning tactile policies and transferring them to real robots. The code first shipped inside the IsaacGymEnvs repository, and today Isaac Lab's visuotactile sensor module is implemented on top of TacSL.

ExampleIsaac Lab's visuotactile sensors offer GelSight R1.5 and GelSight Mini configurations, computing penalty-based normal and shear forces via signed-distance-field queries while also outputting an RGB tactile image, usable directly as an observation for reinforcement or imitation learning.

Also called

TacSL

Related

Tactile Simulation · Vision-Based Tactile Sensor · GelSight · NVIDIA Isaac Lab · Asymmetric Actor-Critic · Sim-to-Real Transfer

Sources

TacSL: A Library for Visuotactile Sensor Simulation and Learning (arXiv 2408.06506)
TacSL 项目主页 (Chinese)
Isaac Lab 文档：Visuo-Tactile Sensor (Chinese)

As of

2026-09