# Changelog All notable changes to the **VLA Handbook** project are documented here, derived directly from the repository's git history. --- ## 2026-06-11 — 行业情报雷达上线 ### Added（行业情报层） - 新增**自动行业雷达**：每日 qwen 联网搜索 + 机器人媒体 RSS → 机械 URL 核验 → append-only 写入 `companies/*.md` 文末「🤖 Moltbot Updates」区与每日 `memory/blog/archives/industry-radar/` 日档；周五综合生成 `companies/industry_mainline.md` 行业判断地图。 - **回填过去 2 个月**积压的融资/产品/IPO 动态（智元合作伙伴大会、Booster/Robot Era/Galbot 融资、Tesla Optimus Gen3、Unitree IPO 过会→提交注册、LimX Luna、Galaxea G0.5 等）。 - `AGENTS.md` 增补该自动化的写入权限矩阵与格式协议；修复 AGENTS.md 损坏的 symlink（曾导致 clone 报 File-name-too-long）。 --- ## 2026-06-10 — 综合层大修（人工 + agent swarm） ### Changed (Theory Mainlines — 全部 10 个总纲两个月来首次同步) - 9 个 `theory/*/ *_mainline.md` 综合了 2026-04-09 之后的 ~109 篇新文章：主题归纳 + 判断更新 + 新旧张力显式记录（每份经对抗审查）。 - `theory/README.md` Explorer's Map 与各分区计数更新（254 → 380 篇）；根 README 统计同步。 ### Fixed (结构与链接) - 71 篇散落在 `theory/` 根目录的文章按生产分类器归位至 10 个主题目录（自 4 月分类链路静默中断以来的堆积）。 - 去重 4 组同论文重复解剖（Hydra/Hyper-DP3、ALAM×2、3DThinkVLA、Potential-Guided FM）；修复 `frontier/rl/` 误嵌套。 - 修复 200+ 条 4 月目录重组遗留的失效内链（paper_index.md 155 条、benchmark/world-model 总纲互链、theory-readme 引用等），机械扫描归零。 - 恢复 4 月重组中被误删的 `theory/benchmark_tracker.md`（自动 SOTA 追踪的协议目标文件）。 - 4 篇明确错放的文章重新归位（OMP、MotionWAM、Mask World Model、GeCO，独立审计发现）。 --- ## 2026-02-15 to 2026-02-27 🆕 ### Added (Theory Deep Dives - Auto Pipeline) - **Agent World Model & Infinity Env**: [`theory/agent_world_model_infinity_synthetic_environments_for_agenti_dissection.md`](./theory/agent_world_model_infinity_synthetic_environments_for_agenti_dissection.md) — 無限合成環境對 VLA 泛化的影響 - **CausalGDP**: [`theory/causalgdp_causality_guided_diffusion_policies_for_reinforcem_dissection.md`](./theory/causalgdp_causality_guided_diffusion_policies_for_reinforcem_dissection.md) — 因果引導的擴散策略 for RL - **MIND Benchmark**: [`theory/mind_benchmarking_memory_consistency_and_action_control_in_w_dissection.md`](./theory/mind_benchmarking_memory_consistency_and_action_control_in_w_dissection.md) — 記憶一致性與動作控制基準 - **OLAF**: [`theory/olaf_world_orienting_latent_actions_for_video_world_modeling_dissection.md`](./theory/olaf_world_orienting_latent_actions_for_video_world_modeling_dissection.md) — 潛在動作定向的視頻世界模型 - **RoboGene**: [`theory/robogene_boosting_vla_pre_training_via_diversity_driven_agen_dissection.md`](./theory/robogene_boosting_vla_pre_training_via_diversity_driven_agen_dissection.md) — 多樣性驅動的 VLA 預訓練增強 - **TACO Benchmark**: [`theory/taco_a_benchmark_for_lossless_and_lossy_codecs_of_heterogene_dissection.md`](./theory/taco_a_benchmark_for_lossless_and_lossy_codecs_of_heterogene_dissection.md) — 異構感知器編解碼基準 - **TwinVLA**: [`theory/twinvla_data_efficient_bimanual_manipulation_with_twin_singl_dissection.md`](./theory/twinvla_data_efficient_bimanual_manipulation_with_twin_singl_dissection.md) — 數據高效的雙臂操作 - **World Action Models**: [`theory/world_action_models_are_zero_shot_policies_dissection.md`](./theory/world_action_models_are_zero_shot_policies_dissection.md) — 世界模型即零樣本策略 - **Scaling Verification**: [`theory/scaling_verification_can_be_more_effective_than_scaling_poli_dissection.md`](./theory/scaling_verification_can_be_more_effective_than_scaling_poli_dissection.md) — 驗證擴展 vs 策略擴展 ### Added (Reports) - **Biweekly 2026-02-26**: [`reports/biweekly/2026-02-26.md`](./reports/biweekly/2026-02-26.md) + [`reflection`](./reports/biweekly/reflection_2026-02-26.md) — VLA 雙周研究報告與預測回顧 - **Weekly 2026-02-20**: [`reports/weekly/2026-02-20.md`](./reports/weekly/2026-02-20.md) — 週度質量評審 + 風向洞察 - **Wind Direction History**: [`reports/weekly/wind-direction-history-2026-02-26.md`](./reports/weekly/wind-direction-history-2026-02-26.md) — 4-layer 框架歷史趨勢 ### Added (Scripts & Automation) - **SCRIPTS.md**: [`scripts/SCRIPTS.md`](./scripts/SCRIPTS.md) — 完整 pipeline 腳本參考（命名規則、各腳本用途、DAG 拓撲） - **vla-trend-snapshot.py**: [`scripts/vla-trend-snapshot.py`](./scripts/vla-trend-snapshot.py) — 週度 VLA 趨勢快照生成器（4-layer 分類） - **backfill-vla-history.py**: [`scripts/backfill-vla-history.py`](./scripts/backfill-vla-history.py) — 歷史數據回填工具 ### Infrastructure Changes - **Watchdog v6**: 15 checks (from 13), 新增 lockfile 防 OOM、Phase 3 孤兒恢復、SOTA/Release 鮮度檢查、AI Deep Dive 預警、磁盤空間監控 - **qwen3.5-plus 全面替換 qwen-max**: `_vla_expert.py`、`rate-vla-daily.py`、expert curation — 質量更好且速度更快 - **VLA 評分超時修復**: Phase 2.5 timeout 180s → 600s，覆蓋 28 篇論文 + arXiv 查詢的正常耗時 - **AI 社交情報修復**: `search_terms_social` 從 28 個發布類詞條替換為 20 個觀點/爭議/病毒類詞條 - **Cron 時間調整**: AI Social 7:30→7:45，VLA SOTA 9:45→9:50，消除 3-way 重疊 - **Calibration timing 修復**: 07:45 → 11:00 Shanghai（VLA pipeline 完成後），vla_release key mismatch 修復 --- ## 2026-01-11 to 2026-01-12 🆕 ### Added - **Spirit-v1.5 深度拆解**: 新增 [`theory/spirit_v1_5_dissection.md`](./theory/spirit_v1_5_dissection.md)，包含： - **核心架构**：Qwen3-VL (大脑) + DiT (小脑) + ODE Euler 积分 (执行)。 - **多样化数据采集 (Diverse Collection)**：深度解析 Spirit AI 为什么认为“干净数据是伟大模型的敌人”。 - **RoboChallenge Table30**：首个超越 π0.5 登顶榜单的代码级复现指南与数据对齐细节（包含不同机器人型号的后处理逻辑）。 - **灵巧手机械学深度专题**: 新增 [`theory/dexterous_hand_mechanics.md`](./theory/dexterous_hand_mechanics.md)： - 机构学基础 (Grubler's Criterion)、四连杆机构、行星减速器、锥齿轮分析。 - 传动学对比 (直驱 vs 线驱 vs 液压/Sanctuary AI Phoenix)。 - 运动学 Jacobian 矩阵与动力学阻抗控制数学表达。 - **具身数据采集概览**: 新增 [`deployment/embodied_data_collection_overview.md`](./deployment/embodied_data_collection_overview.md)： - 覆盖第一视角 POV (EgoScale)、Sim2Real 规模化、真机 RL、专家示教 (GELLO/Manus) 等五大路径对比。 - 深入探讨了“脏数据”、“废数据”与“有效信息密度”的工业界定义。 - **多模态数据同步技术**: 新增 [`deployment/multimodal_data_synchronization.md`](./deployment/multimodal_data_synchronization.md)，解决视觉 (30Hz) 与控制 (1000Hz) 的毫秒级对齐、PTP 协议与硬件触发同步。 - **公司手册更新**: 加入 **Lumos Robot (鹿明机器人)** 专题，解析其“超级数据工厂”定位、鹿明指数与 FastUMI Pro 系统。 - **硬件选型更新**: - 新增 **Sharpa Wave** (DTA 动态触觉阵列)、**LEAP Hand V2 Adv**、**RealerHand (睿尔灵)** 详情。 - **灵巧手典型挑战**：开可乐罐 (指甲利用与杠杆原理)、抓取手机薄片 (桌面碰撞与柔顺性)、重载抓取 (远端关节扭矩要求)。 ### Changed & Optimized - **全站汉化**: 完成所有 140+ 文档的简体中文翻译与本地化，统一技术术语。 - **部署架构重组**: 重新梳理 [`deployment/README.md`](./deployment/README.md) 索引，按硬件选型、感知同步、机械臂控制、灵巧手专题、仿真数采五大板块分类。 - **公式渲染适配**: 全面优化 `math_for_vla.md` 与 `vla_loss_functions_handbook.md`，使用 `$$` 块与空行强制适配 GitHub Markdown 数学渲染。 - **清理不相关内容**: 移除了与机器人技术不相关的 `grade7b_math` 考试目录。 --- ## 2026-01-09 to 2026-01-10 ### Added - **VLA 必备数学基础**: 新增 [`theory/math_for_vla.md`](./theory/math_for_vla.md)，系统整理了从线性代数、空间表示 (SE3) 到扩散模型、流匹配的完整数学链条。 - **VLA 损失函数手册**: 新增 [`theory/vla_loss_functions_handbook.md`](./theory/vla_loss_functions_handbook.md)，包含 NLL、KL 散度、ELBO 等公式大白话翻译与 PyTorch 实现。 - **前馈与反馈控制**: 在 [`theory/robot_control.md`](./theory/robot_control.md) 中增加 Feedforward vs Feedback 专题，辅以“抓取透明水瓶”的工程案例。 - **灵巧手实战案例集**: 将 VisionOS (Webcam 遥操作) 与 Wuji 手的 retargeting 实战整合至 [`deployment/dexterous_hand_applications.md`](./deployment/dexterous_hand_applications.md)。 --- ## 2026-01-06 ### Added - **实战案例：手势控制灵巧手**: 新增 [`deployment/mediapipe_wujihand_project.md`](./deployment/mediapipe_wujihand_project.md)，包含： - MediaPipe + WujiHand 联动架构（WebSocket + USB SDK）。 - **延迟优化实战**：记录从 500ms 降至 50ms 的优化路径（移除软件滤波、启用硬件 LowPass、非阻塞写入）。 - **面试 Q&A 模拟**：针对实时控制、延迟优化、多维映射等核心问题的专业话术建议。 - **Jim Fan 2025 年度复盘**: 新增 [`theory/frontier/jim_fan_2025_robotics_lessons.md`](./theory/frontier/jim_fan_2025_robotics_lessons.md)。 - **触觉不可替代性与视触觉前沿**: - 新增 [`theory/frontier/tactile_irreplaceable.md`](./theory/frontier/tactile_irreplaceable.md)。 - 新增 [`theory/frontier/unitachhand.md`](./theory/frontier/unitachhand.md)（人手→机器人灵巧手策略零样本迁移）。 - 更新 [`theory/tactile_vla.md`](./theory/tactile_vla.md) 的 Q&A 部分，增加「Demo 为何翻车」与「工程 Checklist」。 ### Changed - **理论索引更新**: 在 [`theory/README.md`](./theory/README.md) 中新增「Part 7: 实战案例与部署」板块。 --- ## 2025-12-29 ### Added - **AI Coding Agent Design Deep Dive**: New document [`system-design/ai_coding_agent_design.md`](./system-design/ai_coding_agent_design.md) covering: - User prompt preprocessing (@context, slash commands). - MCP (Model Context Protocol) architecture and tool discovery. - SubAgent implementation and context isolation. - Spec-driven development (OpenSpec interpretation). - **Data Flywheel & Cross-modal Transfer**: New document [`theory/frontier/data_flywheel_and_cross_modal.md`](./theory/frontier/data_flywheel_and_cross_modal.md) analyzing: - Data scarcity solutions in robotics. - Humanoid foundation models using internet video data. - Boundless Intelligence (智在无界) case study. - **Robot Startup Category**: Added [`companies/startups.md`](./companies/startups.md) featuring Boundless Intelligence (智在无界) and CEO Zongqing Lu's team. - **Nature Communications Reward Discovery**: Added [`theory/frontier/reward_discovery_rl.md`](./theory/frontier/reward_discovery_rl.md). ### Changed - **ROS2 Section Reinforcement**: Major updates to [`deployment/ros_and_optimization.md`](./deployment/ros_and_optimization.md): - Added Zero-Copy (Iceoryx) and CycloneDDS tuning. - Added Component Containers and WaitSet real-time executor models. - Added Unicast/Peers-list discovery for distributed deployment. - Added safety certification (Apex.OS) content. - **Robot Dynamics Classification Refinement**: Updated [`theory/robot_dynamics_classification.md`](./theory/robot_dynamics_classification.md) with English terms and algorithm-friendly analogies. --- ## 2025-12-26 to 2025-12-28 ### Added - **OneTwoVLA Adaptation**: Real-world migration guide from 7-DOF Franka to 6-DOF UR5 in `ur5_control_guide.md`. - **Research Frontier Restructuring**: Organized model-specific analyses into `theory/frontier/`. - **Co-training Examples**: Added Robot vs Internet data comparison to π0.5 dissection. ### Changed & Fixed - **GitHub Math Rendering Global Fix**: Optimized LaTeX layout across all 40+ documents, strictly enforcing blank lines around `$$` blocks for correct web display. - **Architecture Diagrams**: Replaced Mermaid with ASCII and enhanced diagrams for GR00T-N1.6, π0, and π0.5. - **Math Deep Dives**: Significantly enhanced explanations for Diffusion Policy and Flow Matching. --- ## 2025-12-21 to 2025-12-25 ### Added - **GR00T-N1.6 Deep Dive**: Detailed analysis of NVIDIA's humanoid foundation model, DiT architecture, and conditioning. - **Tesla Optimus V2 Analysis**: Added hardware analysis of the latest Optimus hand and compared it with the Wuji hand. - **MM-ACT (Unified Token Space)**: Analysis of multi-modality unified token space. - **SGTM (VLA Intrinsic Safety)**: Added chapter on safety and alignment. --- ## 2025-12-15 to 2025-12-18 ### Added - **VLN DualVLN**: Introduced Vision-Language Navigation with Dual-system (Fast/Slow). - **Traditional Action Generation**: Added MSE/GMM explanation to `theory/traditional_action_generation.md`. - **LeetCode for Beginners**: Added oral-style LeetCode training diary to the question bank. ### Changed - **Handbook Restructuring**: Renamed to "VLA Handbook", optimized README with theory-first logic and collapsible sections. - **First Principles Math**: Added dedicated math sections to core theory documents. --- ## 2025-12-10 to 2025-12-13 ### Added - **NeurIPS 2025 Insights**: Decoded top papers from an Embodied AI perspective (Artificial Hivemind, Gated Attention, etc.). - **GELLO Deployment**: Added guide for assembly and UR5 teleoperation using GELLO. - **SaTA (Tactile Awareness)**: Deep analysis added to `tactile_vla.md`. - **Isaac Lab**: Added documentation for the GPU-accelerated simulation framework. ### Fixed - **GELLO Pricing**: Corrected assembled price to ~¥2000 and added Taobao purchase links. --- ## 2025-12-01 to 2025-12-08 ### Added - **Small VLA Research**: Comprehensive analysis of Evo-1, SmolVLA, and Latent Action Learning (UniVLA, MemoryVLA). - **ByteDance GR-RL**: Dissection of the first VLA to achieve 78% shoe-tying success. - **Tactile VLA Module**: Added GelSight, Digit, DM-Tac, and GelStereo sensor analyses. - **VLM Comparison Tables**: Added PaliGemma 3B, SigLIP, and Qwen2-VL comparison for VLA backbones. - **Motion Planning & SLAM**: Added foundational modules for perception and navigation. ### Changed - **README_FUN.md**: Introduced "Robot Growth School" analogy for easier onboarding. - **ASCII Cheat Sheet**: Added visual diagrams for core concepts like LoRA, CoT, and Flash Attention. --- ## 2024-11-25 to 2024-11-30 ### Added - **Product Encyclopedia**: Detailed specs and pricing for 50+ humanoid robots, arms, hands, and sensors. - **Evaluation Protocols**: Math definitions for success rates, A/B testing, and bench details. - **Simulation Guide**: Comprehensive guide on Isaac Sim, MuJoCo, and SAPIEN. ### Changed - **Co-training Refactor**: Extracted Co-training to a dedicated document `theory/co_training.md`. - **Data Pipeline**: Detailed format comparison (RLDS vs LeRobot). --- ## 2024-11-15 (Initial Release) ### Added - **Core VLA Theory**: Transformer vs CNN, RT-1/RT-2 architectures, and action representations. - **Hardware Pricing**: Initial hardware guide and RealSense calibration basics. - **Action Tokens**: Initial research on discrete vs continuous action spaces.