Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

RevoSim: Scalable Multimodal Tactile Simulation for Dexterous Manipulation

Ke Zhang, Lixin Xu, Ziyi Wang, Xiyue Dong, Jie Tan, Lianghu Shao, Xianpeng Wang, Chuanyu Li, Renjing Xu

November 2026

Two simulated Revo3 hands hang a cup on a mug tree while taxel pressure and fingertip RGB, depth and marker panels update.

TL;DR A GPU simulator in Isaac Lab for both kinds of touch sensing on the BrainCo Revo3 hand, 285 piezoresistive taxels plus five visuotactile fingertips. Its fingertip marker-displacement mode fits 4,096 parallel environments on one RTX 5090, and its depth mode uses about 70% less memory than TacMap at 1,024 environments.

What they do

Training tactile policies at scale needs fast simulation of the specific sensors on a hand, but most tactile simulators cover one sensor family or a generic taxel layout. RevoSim simulates both tactile modalities of the BrainCo Revo3 hand in Isaac Lab. These are 285 piezoresistive taxels in eleven pressure regions on the palm and finger pads, plus five visuotactile fingertips. At every step it returns whole-hand pressure, fingertip depth and RGB images (240x320) and the 3D displacement of 100 markers per fingertip. Both modalities are calibrated against real sensor responses, including sensitivity, drift, hysteresis, dead taxels and crosstalk.

How it works

  • Pressure. Each taxel’s penetration into the object’s signed distance field, and how fast that penetration changes, drive a calibrated spring-damper response with its own gain, threshold and saturation. A normal-aware kernel then spreads the response over the curved surface without leaking onto parts that face the other way.
  • Adaptive fingertip depth. A coarse 32x24 penetration map finds the contact. A fixed 25x40 lattice of rays is then cast only around it, and the result is resampled onto the full 240x320 grid, so the ray count follows the contact rather than the image resolution.
  • RGB and markers. GPU-Taxim renders the contact on top of a real no-contact image. Marker motion combines a dilation field from the indentation with a shear term that builds up over time.

Results

  • Depth only: 9.2 GiB at 1,024 environments, about 70% less than TacMap’s 30.6 GiB.
  • Marker displacement: 4,096 environments fit in 31.36 GiB. At that scale FOTS and HydroShear run out of memory, and TacSL already hits the limit at 256.
  • Adding pressure to depth costs little memory. With markers and RGB added as well, capacity tops out at 1,024 environments.
  • The sim-vs-real comparisons of fingertip images, markers and taxel readouts are qualitative. Ball-rotation and grasping policies are shown without success rates.

Limitations

The paper has no policy sim-to-real transfer and no quantitative sim-to-real error, and the memory plots measure capacity, not simulation speed. RevoSim models one specific hand. Deformable objects and richer temporal dynamics are left for future work.

Videos

A simulated Revo3 hand grasps a hammer, and its fingertip depth and marker displacement respond to the contact.
Sensor viewer for both simulated hands, with each fingertip's RGB, markers, depth and tip wrench beside the whole-hand taxel map.

Figures

From Figure 1: Object grasping trained with full-palm tactile and fingertip visuo-tactile observations, showing four ...
From Figure 1: Object grasping trained with full-palm tactile and fingertip visuo-tactile observations, showing four views of the grasp above the palm taxel map and the five fingertip depth images.
From Figure 3: Simulated RGB image of a fingertip indentation, from the qualitative comparison with a real sensor.
From Figure 3: Simulated RGB image of a fingertip indentation, from the qualitative comparison with a real sensor.
From Figure 3: The real fingertip sensor response with tracked markers, for comparison with the simulated image.
From Figure 3: The real fingertip sensor response with tracked markers, for comparison with the simulated image.
From Figure 2: Peak GPU memory for deformation-field simulation against the number of parallel environments, for Revo...
From Figure 2: Peak GPU memory for deformation-field simulation against the number of parallel environments, for RevoSim, TacSL, FOTS and HydroShear. Crosses mark out-of-memory runs.
Abstract

We present RevoSim, a scalable and high-fidelity multimodal tactile simulator for the BrainCo Revo3 hand that jointly models 285 piezoresistive taxels and five visuotactile fingertips. RevoSim efficiently generates synchronized pressure, depth, RGB, and marker-displacement observations; on a single RTX 5090, its depth-conditioned marker-displacement configuration supports up to 4,096 parallel environments. We train in-hand manipulation and grasping policies entirely in simulation, establishing a foundation for future sim-to-real transfer. Project page: https://braincotech.github.io/RevoSim-web/.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.