Wiki Index

这个文件由 Codex 维护。新增、重命名或删除 wiki pages 时必须同步更新。第一部分是研究入口；后半部分是完整 inventory，用于 Obsidian、Quartz 和 deterministic health checks。

从研究问题进入

• 当前总判断：先读 Overview（总览）。它维护当前 wiki 对 embodied AI、robotics simulation、world models、robot foundation models 与 evaluation 的综合判断。
• 问题索引：读 Research Questions。它把当前 wiki 能回答的高价值问题连接到相应 concept/source，而不额外引入复杂 map 层。
• World model 如何影响 robot decision？ 读 World Models for Embodied AI、World Model Evaluation、Latent Dynamics Action Models。
• Inverse dynamics model 怎么从视频学 action？ 读 Inverse Dynamics Models、Latent Dynamics Action Models、Seer、DeFI。
• Robot foundation model 如何处理 heterogeneous data？ 读 Vision-Language-Action Models、Robot Context Conditioning、Latent Dynamics Action Models。
• Simulation benchmark 能证明什么？ 读 Task-Generalist Policy Evaluation、Simulation Sensitivity Analysis、Simulation Reality Gap（仿真现实差距）。
• Simulation infrastructure decisions 为什么重要？ 读 Robotics Simulation Infrastructure，重点看 task/API、asset management、rendering memory/fidelity、visualizer diagnostics 和 pose API 如何影响 RL/evaluation workflow。
• Visual sim-to-real 怎么跨过 reality gap？ 读 Visual Sim-to-Real、VIRAL 和 Simulation Reality Gap（仿真现实差距），重点看 privileged teacher、vision student、domain randomization、hand/camera alignment 与 failure cases。
• OpenUSD 的核心价值是什么？ 读 OpenUSD Scene Composition、OpenUSD 和 Introduction to USD；如果关注 robotics asset authoring，再接 Isaac Sim Asset Structure 3.0 与 Isaac Sim Legacy Asset Structure。
• Isaac Sim Asset Structure 3.0 怎么理解？ 读 Isaac Sim Asset Structure 3.0 和 Asset Structure - Isaac Sim Documentation，重点看 layer role、payload/variant composition 和 engine-specific tuning 隔离。
• Isaac Sim 旧 asset layout 是不是 2.0？ 读 Isaac Sim Legacy Asset Structure 和 Asset Structure - Isaac Sim 4.5 Documentation；当前 source-backed 结论是它应称为 legacy / pre-3.0，而不是 2.0。
• Isaac Sim 的 mujoco.usda 应该放什么？ 读 Isaac Sim mujoco.usda Runtime Semantics 和 Isaac Sim Asset Structure 3.0，重点区分 shared visual/collider asset semantics 与 MuJoCo-only runtime tuning。
• PhysX articulation、position drive、solver 和 joint gains 怎么理解？ 先读 Reduced-Coordinate Articulations 和 Articulations - Omni Physics，再读 Isaac Sim and MuJoCo Physics and Control Notes；PhysX articulation drive 的 PD analogy 和 envelope 已有 official source support，MuJoCo 对比与机械臂 gain grouping 仍是 conversation-derived。
• OBJ、STL、USD、GLB 等 3D 模型格式怎么选？ 读 3D Model Formats Learning Map，注意其中 USD 相关内容已有 wiki source-backed coverage，其他格式仍是 unsourced learning scaffold。
• Contact physics 为什么会影响 learning/control？ 读 Contact Models in Robotics、Contact Complementarity（接触互补）、Contact Solvers（接触求解器）、Differentiable Physics（可微物理）。
• 轮式机器人建模怎么系统学习？ 读 Wheeled Robot Modeling Learning Map、Wheeled Robot Visual Lab、Wheeled Robot Kinematics、Wheeled Mobile Robot Classification、Omnidirectional Wheels、Nonholonomic Mobile Robots 和 Steerable Wheels。

维护入口

• Wiki Log - append-only operation history
• Health check: python3 tools/health.py
• Graph build: uv run python tools/build_graph.py --report
• Publishing preview: npm run wiki:preview
• Production build: npm run wiki:build

Syntheses

• Research Questions - 当前 wiki 支持的高价值研究问题、阅读路径和 evidence boundary
• Isaac Sim mujoco.usda Runtime Semantics - distill mujoco.usda 的 ownership boundary：不是 visual/collision asset 文件，而是 MuJoCo-specific runtime interpretation / tuning overlay
• Isaac Sim and MuJoCo Physics and Control Notes - distill Isaac Sim 官方文档措辞、PhysX/Isaac Sim position drive semantics、stiffness/damping、effort limit、七自由度机械臂 gain scaling、MuJoCo/PhysX solver 差异和参数迁移边界
• 3D Model Formats Learning Map - OBJ、STL、PLY、glTF/GLB、FBX、USD、STEP、URDF/SDF/MJCF 等 3D asset formats 的学习脚手架与 source acquisition plan
• Wheeled Robot Modeling Learning Map - 轮式机器人建模学习脚手架，覆盖 wheel taxonomy、kinematics、swerve allocation、simulation path、failure modes 和 source acquisition plan
• Wheeled Robot Visual Lab - 内嵌 academic-style 交互图，把 wheel-level constraints、contact-point velocity 和 chassis matrix rows 放在同一张平面图中复习

Sources

OpenUSD And Asset Infrastructure

• Introduction to USD - OpenUSD 官方 introduction，解释 scene description、Stage/Layer/Prim data model、schemas、composition arcs、Hydra、extension points 与边界条件

World Models

• A Comprehensive Survey on World Models for Embodied AI - embodied AI world models 的 POMDP/ELBO formalization、三轴 taxonomy、datasets/metrics 与 open challenges
• AwesomeWorldModels - survey companion GitHub repository，按 taxonomy 维护 world model bibliography

Robot Foundation Models

• Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation - Seer/PIDM 框架，用 foresight token 和 inverse dynamics action token end-to-end 连接 vision prediction 与 action prediction
• Disentangled Robot Learning via Separate Forward and Inverse Dynamics Pretraining - DeFI 框架，把 visual forward dynamics 和 inverse dynamics 分开预训练，再耦合微调到 robot actions
• LDA-1B: Scaling Latent Dynamics Action Model via Universal Embodied Data Ingestion - dynamics-centric robot foundation model，用 DINO latent、MM-DiT 和 EI-30K 做 role-aware heterogeneous data ingestion
• π0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities - Physical Intelligence 的 steerable VLA model，强调 context conditioning、subgoal images、metadata 与 compositional robot generalization

Simulation And Evaluation

• VIRAL: Visual Sim-to-Real at Scale for Humanoid Loco-Manipulation - project page for RGB-based humanoid loco-manipulation，覆盖 privileged RL teacher、vision student distillation、visual randomization、finger SysID、FOV alignment、generalization videos 和 failure cases
• Robotics Simulation Infrastructure - Stone Tao 的 simulation infrastructure blog，覆盖 task/API、asset management、physics/rendering、visualizer、ML integration、rendering memory/fidelity 和 pose API trade-offs
• RoboLab - RoboLab official implementation repository，包含 Isaac Lab task library、predicate/subtask system、policy clients 与 analysis tooling
• RoboLab: A High-Fidelity Simulation Benchmark for Analysis of Task Generalist Policies - NVIDIA 的 high-fidelity simulation benchmark，用 RoboLab-120、language variants 与 sensitivity analysis 评测 task-generalist robot policies
• Asset Structure - Isaac Sim 4.5 Documentation - Isaac Sim 4.5 docs 中的 legacy / pre-3.0 asset layout，覆盖 asset_base.usd、parts.usd、asset_sim_optimized.usd、feature layers 和 final asset.usd
• Asset Structure - Isaac Sim Documentation - Isaac Sim 6.0 EDR docs 中的 USD Asset Structure 3.0 guidance，说明 robot assets 的 geometry/material/instance/physics/runtime/schema/feature layer organization
• Articulations - Omni Physics - NVIDIA Omni Physics docs 中的 PhysX reduced-coordinate articulation guidance，覆盖 root/topology、JointStateAPI、drive performance envelope、joint friction、closed loops、mimic joints 和 tendons

Contact Physics

• Contact Models in Robotics: a Comparative Analysis - robotic contact models 与 contact solvers 的 comparative survey 和 benchmark

Wheeled Robot Modeling

• Modern Robotics Chapter 13: Wheeled Mobile Robots - Lynch 和 Park 的 wheeled mobile robot chapter，覆盖 planar chassis twist、omni/mecanum kinematics、nonholonomic canonical model、Lie bracket controllability、odometry 和 mobile manipulation
• Structural Properties and Classification of Kinematic and Dynamic Models of Wheeled Mobile Robots - Campion、Bastin 和 D’Andrea-Novel 的 WMR taxonomy source，提出 degree of mobility、degree of steerability 和五类 nondegenerate WMR

Concepts

OpenUSD And Asset Infrastructure

• OpenUSD Scene Composition - 把 OpenUSD 的 Stage/Layer/Prim data model、schemas、composition arcs、Hydra 与 Isaac Sim robotics asset layering 连接起来的学习页

World Models

• World Models for Embodied AI - action-conditioned latent simulators 的 POMDP/ELBO 机制与 practical failure modes
• World Model Taxonomy - Functionality、Temporal Modeling、Spatial Representation 三轴分类
• World Model Evaluation - world model metrics 从 pixel fidelity 到 state understanding 与 task performance 的评估层次
• Latent Dynamics Action Models - 用 DINO latent、diffusion action chunks 与 multi-task objective routing 学习 robot interaction dynamics
• Inverse Dynamics Models - 从 current/future visual transition 推断 action 或 latent action 的模型，包含 DeFI/GIDM 的 self-supervised video pretraining 机制

Robot Foundation Models

• Vision-Language-Action Models - 从 observation history 和 context 预测 robot action chunks 的 VLA policy formalism
• Robot Context Conditioning - 用 task/subtask language、metadata、subgoal images 与 control mode disambiguate heterogeneous robot data
• Compositional Generalization in Robotics - robot policies 如何把 seen skills、objects、embodiments 与 instructions 重新组合到 unseen tasks

Simulation And Evaluation

• Task-Generalist Policy Evaluation - 用 task libraries、language variants、predicates、subtask scoring 和 diagnostics 评估泛化 robot policies
• Robotics Simulation Infrastructure - 把 simulator framework 拆成 task/API、asset management、physics/rendering、visualizer 和 ML loop 的 design decisions
• Simulation Sensitivity Analysis - 用 controlled perturbations 与 NPE/MNPE posterior 找出影响 robot policy success 的环境参数
• Simulation Reality Gap（仿真现实差距） - contact approximations、learned dynamics 和 policy context 到 sim-to-real mismatch 的 causal flow
• Visual Sim-to-Real - VIRAL-style visual sim-to-real 的 teacher-student formulation、domain randomization、real-to-sim hand/camera alignment 与 deployment failure modes
• Isaac Sim Legacy Asset Structure - Isaac Sim 4.5 legacy / pre-3.0 asset layout，明确旧 layout 没有 source-backed 2.0 命名，并与 Asset Structure 3.0 对照
• Isaac Sim Asset Structure 3.0 - 用 architecture diagrams、USD layers、payloads、references 和 variants 组织 Isaac Sim robot assets，并隔离 neutral physics 与 MuJoCo/PhysX-specific tuning
• Reduced-Coordinate Articulations - PhysX / Omni Physics articulation 的 mechanism-level page，覆盖 reduced coordinates、root selection、drive envelope、mimic/tendon constraints 和 solver failure modes

Contact Physics

• Contact Models in Robotics - simulator contact-law choices 作为 modeling assumptions，并包含 contact pipeline 图
• Contact Complementarity（接触互补） - rigid contact、friction cone、NCP/LCP/CCP 与 residual intuition
• Contact Solvers（接触求解器） - 计算 contact forces 与 impulses 的 numerical methods，以及 solver taxonomy
• Differentiable Physics（可微物理） - simulator gradients、contact artifacts 与 gradient contamination

Wheeled Robot Modeling

• Wheeled Robot Kinematics - wheel speeds、steering angles、body twist、rolling/no-slip constraints 和 $u=H(0)V_b$ 的统一建模入口
• Wheeled Mobile Robot Classification - Campion taxonomy 中的 ${\delta }_m$、${\delta }_s$、${\delta }_M$ 与五类 WMR
• Omnidirectional Wheels - omniwheel / mecanum 的 rank condition、feasible twist polyhedron、roller contact failure modes 和仿真实践含义
• Nonholonomic Mobile Robots - unicycle、diff-drive、car-like robot 的 Pfaffian constraint、canonical model 和 Lie bracket controllability
• Steerable Wheels - centered steerable wheels、off-centered caster wheels、swerve-style modules 和 steering DOF failure modes
• Mobile Robot Odometry - 用 wheel encoder increments 估计 chassis pose 的 $H^{†}$ / SE(2) integration 方法与 drift sources

Entities

Models, Datasets, And Benchmarks

• DeFI - decoupled forward/inverse dynamics pretraining framework，用 GFDM、GIDM 和 action adapter 从 videos 到 robot commands
• EI-30K - LDA-1B source 构建的 30k+ hour heterogeneous embodied interaction dataset
• LDA-1B - dynamics-centric robot foundation model，统一 policy、latent dynamics 和 visual forecasting
• Seer - end-to-end PIDM model，用 [FRS] foresight token 和 [INV] action token 做 robot manipulation policy learning
• π0.7 - steerable generalist VLA model，使用 language、metadata、subgoal images 和 control mode conditioning
• RoboLab - high-fidelity simulation benchmark/platform for task-generalist robot policy evaluation
• VIRAL - visual sim-to-real framework for humanoid loco-manipulation，使用 privileged teacher、vision student distillation、visual randomization 和 real-to-sim alignment
• ContactBench - source 中的 unified C++ contact-model benchmark framework

Organizations And Repositories

• AwesomeWorldModels - survey companion bibliography repository
• Galbot - LDA-1B source 中的作者机构与 Galbot G1 real-world evaluation platform context
• NVIDIA - RoboLab paper/project/repo、VIRAL-linked NVlabs code repository 与 Isaac Sim Asset Structure docs 的发布机构
• Pixar - OpenUSD official docs 的 copyright holder 与 USD production-pipeline heritage context
• Physical Intelligence - π0.7 source 中的 robot foundation model research organization

Scene Description Platforms

• OpenUSD - Universal Scene Description / OpenUSD scene-description platform；当前 coverage 聚焦 official Introduction 与 Isaac Sim asset-structure usage

Simulation Tools

• Isaac Sim - NVIDIA robotics simulation stack；当前 source-backed coverage 包括 legacy / pre-3.0 Asset Structure、Isaac Sim 6.0 EDR Asset Structure 3.0 与 Omni Physics articulation semantics
• PhysX - NVIDIA physics runtime / SDK family；当前 wiki coverage 聚焦 Omni Physics articulations source 中的 reduced-coordinate mechanisms、drive envelope、mimic joints 和 tendons
• MuJoCo - source 中用于讨论 contact regularization tradeoffs 的 robotics physics engine；Isaac Sim docs 中也作为 engine-specific asset layer 出现
• ManiSkill - Stone Tao article 中用于说明 Python API、batched rendering performance 和 Pose abstraction 的 robotics simulation framework
• RaiSim - source 中用于讨论 quadruped transfer 与 per-contact handling 的 robotics simulator