Skip to content
Bud's Wiki

Project SuperDex — A unified simulation platform for dexterous manipulation

TL;DR

Section titled “TL;DR”

Project SuperDex is a Meta open-source simulation platform for dexterous manipulation research, built around a purpose-built contact-first physics engine and packaged as four layers: SuperDex Physics (multi-physics solver for rigid + soft articulations, non-convex collision, rods/tendons, shells/cloth, tactile sensors), SuperDex Robotics (declarative .superdex_bot robot definitions, URDF import, controllers, sensors), SuperDex Studio (authoring / bot composition / mesh + CAD import), and SuperDex Lab (Gymnasium-compatible RL interface with Ray/RLlib). The distinguishing bet is a single nonlinear-optimization solver core underpinning both forward dynamics and constraint-aware inverse kinematics, with VR teleoperation via Meta Quest 3 as the demonstration-data pipeline. SuperDex Lab ships as “early preview” — the physics + authoring + robotics stack is the load-bearing release, RL API is nascent.

Key claims

Section titled “Key claims”
  • Unified solver spans rigid bodies, soft bodies, rods & tendons, and shells & cloth in one physics engine [Overview §Multi-Physics Simulation].
  • Non-convex collision produces accurate contact force distributions for arbitrary geometries including deforming objects [SuperDex Physics §Non-Convex Collision].
  • The engine claims numerical stability without the restrictive time-step limits of explicit / semi-implicit methods [SuperDex Physics §Numerical Stability].
  • Constraint-aware inverse kinematics is built on the same nonlinear optimization core as forward dynamics, so IK poses satisfy collision, end-effector, and trajectory constraints under a shared solver [SuperDex Physics §Inverse Kinematics].
  • Robot description is a human-readable .superdex_bot file that extends physics articulations with robotics-workflow fields and is the source of truth from Studio through simulation [SuperDex Robotics §Declarative Bots].
  • URDF imports work directly at runtime; Studio can upgrade a URDF into a native bot with production-quality collision meshes [SuperDex Robotics §BYO Bots].
  • Controllers / sensors / actuators sit under one extensible framework with built-in operational-space and joint-space PD controllers [SuperDex Robotics §Controllers, Sensors, Actuators].
  • Every Studio editor is a live SuperDex Physics scene — simulation is available the instant an asset opens for on-the-fly inspection of contact, inertial, and joint behavior [SuperDex Studio §Simulate].
  • Studio ships a “bot mod pipeline” for swapping OEM fingertips with custom sensors and bolting hands onto arms [SuperDex Studio §Compose, Edit, and Combine Bots].
  • SolidWorks and NX CAD plugins export assets and robots directly with simulation-ready meshing [SuperDex Studio §SuperDex CAD Exporter].
  • SuperDex Lab exposes a Gymnasium-compatible RL framework and an off-the-shelf Ray/RLlib integration for training controllers on top of the physics engine [SuperDex Lab §SuperDex Gym, §RLlib Training].
  • VR teleoperation via Meta Quest 3 drives simulated robots in real time to collect demonstration datasets at reduced cost vs. real-world collection [Gallery, Overview §02 Teleoperate].

Method

Section titled “Method”

The platform is layered vertically. SuperDex Physics is the load-bearing artifact: a contact-first nonlinear-optimization solver that handles rigid + soft articulations, non-convex collision, rods/tendons, shells/cloth, tactile sensors, and soft contact under one framework. Numerical stability is claimed to avoid the time-step limits typical of explicit / semi-implicit integrators, and the same solver core is reused for constraint-aware inverse kinematics. SuperDex Robotics sits above physics with a declarative .superdex_bot schema (links + joints + mesh references + robotics-workflow metadata), a library of ready-to-load arms / hands / sensors, and runtime URDF import as an escape hatch for the wider ecosystem. SuperDex Studio is the authoring tool: compose bots from vetted components, mesh editing / re-mesh / SDF-baking modifier stacks, live physics preview per editor, and SolidWorks + NX CAD-exporter plugins. SuperDex Lab is the RL layer — Gymnasium-compatible environments with Ray/RLlib training, explicitly labeled “early preview” and promised a partially-observable-MDP generalization for RL / system identification / MPC. The teleoperation pipeline connects Meta Quest 3 headsets to simulated hands for demonstration collection.

Results

Section titled “Results”

Landing page — no benchmarks, ablations, throughput numbers, or sim-to-real transfer results are reported. The gallery shows eight interactive-VR-teleop clips (fine-grained dexterous manipulation among them) running the physics in real time. No quantitative comparison to MuJoCo / Isaac Lab / Genesis / SAPIEN is offered. No paper or arxiv preprint is linked from the site; the roadmap and credits pages 404 at filing time.

Why it’s interesting

Section titled “Why it’s interesting”

SuperDex is Meta’s entry into the dexterous-manipulation simulator space that Genesis World — Simulation platform for general-purpose robotics & embodied AI learning (Apache-2.0, independent) and NVIDIA’s Isaac Lab currently anchor — the four-layer packaging (Physics / Robotics / Studio / Lab) more closely matches Genesis’s “sim engine + editor + RL” scope than Isaac’s “GPU physics + Gym wrapper” scope, but the CAD-exporter and Studio-authored bot-mod pipeline are unique commercial-grade features neither ships. It complements Tactile Genesis: Exploring Tactile Sensors at Scale for Learning Dexterous Tasks‘s “eight tactile abstractions under one interface” bet with its own tactile-sensor claim inside a unified solver, and provides a plausible substrate for the Human-to-Robot Retargeting and Tactile sensing for manipulation cluster’s VR-teleop → sim-demos → dexterous-policy pipeline that Simulation Pre-training for Dexterity (SPD) SPD ran on MuJoCo. Trust caveat: no quantitative claims, no code links surfaced from the landing page, no third-party benchmarks — everything below the Physics layer is a marketing pitch until the release drops.

See also

Section titled “See also”