facebookresearch/project_superdex

SuperDex brings together a purpose-built physics engine, robotics authoring tools, and a scalable reinforcement learning interface in a unified simulation platform, with VR-based teleoperation and additional capabilities planned for future releases.

What it solves

Project SuperDex provides a unified simulation platform specifically designed for dexterous manipulation research. It addresses the need for high-fidelity physics, robot authoring tools, and scalable interfaces for training AI policies to handle complex tactile tasks, such as manipulating threaded screws, soft sponges, or rope braids.

How it works

The platform is composed of four primary building blocks:

  • SuperDex Physics: A contact-first physics engine optimized for tactile manipulation and accurate sensing.
  • SuperDex Robotics: An SDK for defining robots, controllers, sensors, and actuators, which are aggregated into simulation configurations.
  • SuperDex Studio: A desktop GUI application used to create and validate simulation assets, including robots, meshes, and scenes.
  • SuperDex Lab: A simulation harness that provides a Gymnasium-style API to facilitate reinforcement learning (RL), Model Predictive Control (MPC), and system identification.

Who it’s for

It is designed for robotics researchers and developers focusing on dexterous manipulation and the development of AI-driven control policies.

Highlights

  • Contact-First Physics: Purpose-built engine for stable contact and tactile sensing.
  • RL Integration: Gymnasium-style API for seamless policy development.
  • Asset Authoring: Dedicated GUI (SuperDex Studio) for converting CAD data into simulation assets.
  • Cross-Platform: Supports Linux, Windows, and macOS.
  • Future Teleoperation: Planned VR-based teleoperation (SuperDex Teleop) for Quest 3 to enable low-latency virtual control.

Related

  • Project
    Genesis-Embodied-AI/genesis-world

    A simulation platform for physical AI and robotics that combines a multi-physics engine, photo-realistic rendering, and a cross-platform compiler for scalable research.

  • Project
    google-deepmind/mujoco

    A high-performance physics engine for the fast and accurate simulation of articulated structures, widely used in robotics and machine learning research.

  • Project
    syswonder/robonix

    An operating system for embodied AI that decouples AI models from robot hardware, enabling a "train once, deploy on any robot" workflow across diverse platforms.

  • Project
    RoboDojo-Benchmark/RoboDojo

    A unified sim-and-real benchmark for evaluating generalist robot manipulation policies across 60 tasks and five capability dimensions.

  • Project
    ihmcrobotics/ihmc-open-robotics-software

    A comprehensive software stack for the control and simulation of humanoid robots, providing tools for dynamics, perception, and real-time communication.