SuperDex Atlas#
Learning home · SuperDex contents
Course text is primarily Chinese, with English terminology and summaries. The original content is preserved.
Understand SuperDex through native APIs, physics concepts and versioned source code.
中文 · Introductory guide · Curriculum · Source map · Versions · Roadmap · Project tracker · Series home
Part of Sim Atlas, an independent community learning series with two complete planned tracks: applications (modeling, control, robotics, sensing and data) and principles/source (dynamics, contact, solvers, integration and extensions).
The initial guide, pinned source map, and E1 modeling, coordinates, state and time chapter are available in Chinese. E1 covers native frames/inertia, URDF import limits, state dimensions and ownership, snapshots, stepping and integration foundations; it includes answered reading exercises and an unexecuted native API snippet. See validation and limitations. A1/A2 are documented; B0/B4 foundations are documented, with dynamics assembly and numerical solving developed further in E3 below.
The E2 control, robotics and task-interface chapter now covers A3/A5/A7: native DoF mapping, JSC/OSC, implicit pose tracking, transmissions, FK/IK, limits, callbacks and application-owned task scheduling. Its native command snippet remains unexecuted; see E2 validation. Batch and data interfaces are developed in E5 below; detailed engine extensions are developed in E6 below.
The E3 contact, solvers and force-observation chapter documents A4 and B1–B5, extending B0 with dynamics assembly. It explains directional contact geometry, field-specific geometric-mean material rules and static-collider exceptions, penalty and regularized friction equations, Backward Euler assumptions, island/stage execution, nonlinear and linear solvers, convergence modes, and force/query timing. It includes 16 answered exercises and an unexecuted native reader; see E3 validation. This fixed source does not establish the binary or solver configuration of historical DexLab runs. Deformable/differentiable extensions are developed in E6 below.
The E4 sensors, rendering and visualization chapter documents A6: CameraSensor metadata versus image production, distance/ray/force/tactile boundaries, RGB/RGBA and missing depth/segmentation contracts, native projection and distortion, frame delays, and GUI/headless ownership. It traces Polyscope, Mochi renderer, UnrealCV and Studio separately, including a pinned Polyscope 2.5.0 reference for its actual RGBA screenshot contract. Sixteen answered exercises, a native metadata reader and a render-readback snippet are source/syntax checked only; no native imports, rendering or simulation were performed. See E4 validation. UE/plugin server identity and depth/mask semantics remain explicit gaps, not invented native capabilities.
The E5 batching, learning and data chapter covers A8/A9 and B6: independent versus snapshot-shared scenes, process/serial hybrid vectorization, explicit per-environment seeds, NEXT_STEP transition masks, Dict flatten order, CPU shared buffers and CUDA solver transfers, native differentiation versus external learning, randomization, HDF events, snapshots and policy checkpoints. Gymnasium 1.3.0 is pinned separately to explain the external seed/autoreset contract. Twenty answered exercises, a minimal native vector snippet and a cached-step reader are source/syntax checked only; no native imports, stepping, rendering, training or performance runs were performed. See E5 validation. Adjoint internals and specialized engine extensions are developed in E6 below.
The E6 specialized actors, differentiation and extension-boundaries chapter deepens B4 with differentiability internals, extends B6, and provides B7 source traces. It explains volumetric FEM, shell/rod units and state, skinning, ROM/hyper-reduction/Deep Flow gates, missing Python ROM fields, implicit adjoints with finite-difference Hessian-vector products, Krylov/Newton branches, Lie gradients, contact-force VJPs and the rigid-only full-Jacobian implementation. Twenty answered exercises and a one-step native backward example are source/syntax checked only; no numerical backward solve or native execution occurred. See E6 validation. A0 installation coverage and the final two-track course audit remain E7 work.
This phase prioritizes understanding engine subsystems. Minimal snippets support explanation and are explicitly marked when unexecuted. No new simulation campaigns, benchmarks, training or scoring are included; later experimental material will reuse DexLab with its original version and workload boundaries.