Full-Hand Tactile Simulation & Algorithm Demos

Explore Revo 3 simulation, learning, and data workflows, from full-hand tactile simulation to real-robot manipulation.

Full-Hand Tactile Simulation

RevoSim provides bimanual multimodal tactile simulation based on NVIDIA Isaac Sim, including fingertip visuotactile images, depth, marker displacement, six-axis contact force/torque, and whole-hand piezoresistive responses. Read data through its Python API or run interactive demos.

RevoSim · Full-Hand Tactile Simulation

Explore bimanual tactile simulation, sensor visualization, and the project overview.

View Project & Demos ↗

Source & Installation

Find environment requirements, the Python API, sensor data formats, and runnable examples.

Open GitHub Repository ↗

Algorithm Demos

The collection covers teleoperation, reinforcement learning, imitation learning, touch, grasp generation, and egocentric data processing in real-robot and simulation scenarios. Each demo describes its setting and development stage. Refer to the individual projects for installation and reproduction instructions.

Browse Revo 3 Algorithm Demos ↗

Retargeting (Real Robot)

RL Rubik’s Cube Solving (Simulation)

Simulation, Data & Algorithm Workflows

The diagrams below are adapted from the Revo 3 demo collection. “Coming soon” marks planned work; “License required” marks restricted access. Models and methods illustrate workflows. Check each project's public documentation for availability.

THE REVO3 ECOSYSTEM

From simulation to real robots

Simulation and Hardware infra

01

Sim infra

Real-to-sim-to-real, SysID

RevoSim ↗

Full-hand multimodal tactile simulation

Simulation foundation

RevoLab ↗

Reinforcement learning tasks and demos

RevoBench Coming soon

Tactile evaluation on a standardized, deterministic platform

02

Robot infra

Post-training infra and pipeline

1. Teleoperation ↗

   • Hand Retargeting
   • Arm IK
   • Revo3 SDK ↗

2. Data collection

   License required
   • frame alignment
   • UI
   • storage

3. Training

   Coming soon
   • ACT
   • pi05
   • FastWAM
   • LingbotVA 2.0

4. Deployment

   Coming soon
   • protocol
   • RTC

Data and Model infra

03

Data infra

Egocentric data pipeline

1. Egocentric data

2. Hand tracking

3. Retargeting

   To robot configuration

4. Pseudo-action data

5. Dense language annotation

6. Data storage

04

Model infra

Hierarchical model stack

RevoAgent

Coming soon

System 2

Inputs

Language instructions

Robot image observations

• Grounded task planning
• Tool & skill orchestration
• Memory & in-context learning
• Verify & replan

RevoNet

Coming soon

System 1

Inputs

Robot image

Proprioception

1. Pre-trainingDiverse open-source models
2. Mid-trainingEgo + RevoHuman
3. Post-trainingReal-world data

System 0

Inputs

Joint angles

Force and Tactile

• High-frequency tactile expert
• RL foundation model
• HIL-RL
• DAgger

• Impedance Control
• Admittance Control
• Position Control
• RTC

Tactile infra

05

Tactile infra

Tactile data conventions & sources

RevoTac

Tactile data convention

Visuotactile · four modalities

• RGB
• Depth
• Marker displacement
• 6D force / torque

Whole-hand piezoresistive touch

• Normalized piezoresistive response
• Spatial distribution map

Simulation data

• Optimization-based generation
• RL rollouts
• Teleoperated grasps
• GPT-generated data

Real-world data

• Real-robot teleoperation
• Deployed policy rollouts
• Human-in-the-loop interventions

RevoSim piezoresistive outputs are normalized responses that may exceed 1, rather than measurements in Pa, N, or resistance. See the repository data reference for six-axis contact force/torque definitions.

Continue development: RevoLab RL Guide · Revo 3 SDK · Ecosystem & Applications