RoboParts 机器人零部件兼容性

RoboParts — 具身系统的组合校验层（开源）

RoboParts — vendor-neutral composition checker for embodied systems. 判定零件能否装配 / 共装 / 互换，输出可核验的三态结论（能 / 不能 / 不知道），并生成可 3D 打印的转接件。我们不卖零件，所以没有理由骗你选一个。

1. Project Overview

• Name: RoboParts — 具身系统的组合校验层（Vendor-neutral composition checker for embodied systems）

• Positioning: 面向开源机器人构建者的组合校验层。判定零件能否装配/共装/互换，输出可核验的三态结论（能/不能/不知道），并生成可打印转接件。开源、中立、可追溯。

• English positioning: Vendor-neutral composition checker for embodied systems — decide whether parts can be assembled / co-mounted / interchanged, returning verifiable three-state verdicts (compatible / incompatible / unknown) and printing transfer adapters. International anchoring: ISO 9409-1 (flange geometry) + ISO 22166-1 (modular frameworks for service robots, Germany-led, revision ISO/AWI 22166-1 FDIS 2026-04-05) + GB/T 20255807-T-604 (draft, domestic modular information model).

• Why not a "part database": 数据只是载体，真正不可复制的是判定引擎（三态 fail-closed compose）+ 三条代数律（可靠性 / 证据单调性 / 不可判定边界，经性质测试与变异测试验证）+ 三条研究假说（其中 T2 已被证伪，见 NR-001；T3 为强假说，样本 56 对）+ 78 项 CI 闸门 + 溯源纪律。

• Domain: roboparts.cc

• Live URL / 线上地址: https://roboparts.cc

• 线上地址（正式入口，外链一律指向此域）: https://roboparts.cc

• Preview domain / 预览域（Cloudflare Pages 默认域，非正式入口，勿对外引用）: https://robotparts-924.pages.dev

• Data volume / 数据量: 868 entities / 868 实体 (659 physical parts / 101 interface specs / 81 AI model software / 10 organizations / 17 market intelligence);Mechanical interface declaration rate 16.2% (81/500);Mechanical interface decidable rate 6.57% (57/868, standard+flange both non-empty);Open-source components 325

• Last updated / 最后更新: 2026-10-19

2. Core Goals

1. Modular accessories — standardized module parameter library, supporting filtering and combination by degrees of freedom / torque / size / protocol
2. Interface standardization — align with the national standard General Technical Requirements for Humanoid Robot Modularity, build a compatibility database
3. Protocol standardization — EtherCAT/CANopen/ROS2/MQTT protocol stack comparison, real-time performance benchmarks, cross-protocol bridging
4. Large-model industrialization — deployment hardware requirements and inference performance comparison for VLA models (RT-2/OpenVLA/π0/GR00T N1.7/SmolVLA/π0.5/τ(0)-VLA/InternVLA-A1.5/Gemini Robotics 2)
5. User data integration — HTTPS transport + KV at-rest encryption + API-key isolation for selection plans / BOM / design files, protecting user habits and privacy

3. Three Principles

• Automation — intelligent selection engine, automatic BOM generation, AI compatibility matching, one-click export of procurement lists
• Ecosystem — connect vendors / developers / users into a closed loop
• Profitability — API data subscription, selection-tool SaaS, enterprise customization (no transaction commission: once commission is taken, judgment is no longer neutral)

4. Core Features

• 🔧 Intelligent selection engine — five-dimension scoring (torque / speed / accuracy / weight / cost) multi-factor selection + multi-result comparison table
• 📋 Data quality labeling — every entity labeled with source / confidence / last_verified
• ✅ Standard conformance — ISO 22166-1 (modularity) / ISO 9409-1 (flanges) / GB modular general tech / GB/T 20255807-T-604 (draft) / ISO 8373 / IEC 61508 / ROS2 compatibility status
• 🧬 Bionic categories — SEA series-elastic actuators, flexible actuators, bionic spine, dexterous hands, artificial muscles
• 🔗 Compatibility matrix — electrical / mechanical / protocol / software four-dimensional compatibility detection

5. Data Categories

• actuators: 222 entries — Actuators (motors, harmonic reducers, planetary roller screws, frameless torque motors, drivers, joint modules, dexterous hands, tendon-driven hands, open-source force-controlled joints, SEA, flexible actuators)
• chips: 108 entries — Chips (AI chips, edge inference accelerators, MCU, FPGA, communication chips)
• platforms: 106 entries — Robot platforms (including open-source reproducible complete robots)
• sensors: 95 entries — Sensors (vision cameras, six-axis force/torque sensors, joint torque sensors, tactile sensors, magnetic e-skin, LiDAR, IMU)
• protocols: 64 entries — Communication protocols (EtherCAT, CANopen, ROS2, MQTT, etc.)
• data_acquisition: 46 entries — Data acquisition devices (teleoperation, exoskeleton capture, motion capture, data gloves, open-source embodied dataset platforms)
• robot_ai_models: 46 entries — Robot AI models (VLA models, world models, robot foundation models)
• interfaces: 44 entries — Interface standards (flanges, buses, connector standard texts)
• llms: 42 entries — Large models (VLA models, robot foundation models)
• grippers: 26 entries — Grippers and end effectors
• flexible_actuators: 22 entries — Flexible actuators (artificial muscles, flexible actuators, bionic spine)
• bionic_mechanisms: 17 entries — Bionic mechanisms (bionic joints, bionic actuators, bionic sensors)
• reducers: 14 entries — Reducers (harmonic, planetary, RV)
• controllers: 4 entries — Controllers
• structural: 3 entries — Structural parts
• cables: 2 entries — Cables
• connectors: 2 entries — Connectors
• pcb: 2 entries — PCB
• power: 2 entries — Power supplies
• integrated_joints: 1 entries — Integrated joint modules

The counts in this section are computed live by scripts/inject_readme_stats.py from api/entities.json and are not hand-written. Previously they were stuck at an early-August snapshot (actuators wrote 217 vs. actual 220, and 10 whole categories were missing) for 7 weeks unnoticed — because the total is watched by a gate, but the subcategories were not.

6. Latest Updates (2026-08-05)

• Open-source hardware upstream gap filled (+10, all Tier A verified) — located structural gaps by corpus-gap probing (rather than re-dumping prioritized categories): tendon-driven transmission had only 1 entry, e-skin only 2, open-source force-controlled joint modules missing. Added dexterous hands LEAP Hand (CMU, 16DoF direct-drive) / RUKA Hand (NYU, tendon-driven), open-source force-controlled joint ODRI Actuator (quasi-direct-drive, torque-sensor-free), drivers mjbots moteus (CAN-FD) and VESC, magnetic e-skin AnySkin / ReSkin (recalibration-free, replaceable), AgiBot World embodied dataset platform (corresponds to direction-202608 P0 "AgiBot supply chain"), open-source complete robots Open Duck Mini / Reachy 2. All 10 entries were elevated to Tier A via real HTTP verification (200 + page hit) by scripts/verify_vendor_sources.py, traceability rate 54.42%→55.20%, total entities 577→587
• Embodied data acquisition / edge compute / domestic sensors expanded on three fronts (+33) — filled P1/P2 gaps from the monthly direction not yet delivered: embodied data acquisition devices 15→26 entries (Mobile ALOHA, Open-TeleVision, AirExo-2, Bunny-VisionPro, ACE, DOGlove, HumanPlus, FastUMI, Manus, Rokoko, Noitom), VLA/edge inference accelerators +12 entries (Axelera Metis, Axera AX650N/AX8850, Black Sesame A2000, Sophgo BM1684X, Cambricon MLU370-S4, Rockchip RK3576, SiMa.ai Modalix, Ambarella N1-655, TI AM69A, DEEPX DX-M1, Houmo M30), domestic-substitute sensors +10 entries (Keli, Anpei, Fenyix, AVIC Electronics, SINOXCG, Huayi, Orbbec Gemini 335, Tyzuum FM851, Hesai FT120, RoboSense AC1), total entities 544→577
• Domestic-substitution index — added the import_substitution_for field, which can directly answer "who is the domestic substitute for this overseas component"
• Humanoid robot supply-chain entity expansion (+51) — first systematic coverage of harmonic reducers (Leaderdrive LCS/LCD/CSG, Harmonic Drive, Laifual, Tongchuan, SZC), planetary roller screws (Zhejiang XCC, Hengli Hydraulic, Best, Qinchuan Machine Tool, Beite, Dianzhi, Shuanglin, Rollvis, GSA, Ewellix, SKF, Rexroth), frameless torque motors (Kinco FMK/FMC, Leadshine FM1/FM2, HCFA Hu-MDB, Inovance MX/TMB, Chuanzhi, Veichi, Wolong, HAN'S, Qianghe, Moons'), six-axis force / joint torque sensors (SRI M35XX/C025XX/C075XX/M221X/M37XX, KWR-N, Hypersen, XJCSENSOR, SHENYUANSHENG, ME K3D, Chuanzhi), integrated joint modules (Leaderdrive, Chuanzhi, Kinco, Leadshine, Tuopu, Sanhua, Zhaowei, Topband), total entities 493→544, over-achieving the August "500+" goal
• Supply-chain field system — added the structured supply_chain field (tier / customers / capacity / domestic_share), supporting retrieval by Tier1/Tier2 levels and localization rate

7. Historical Updates (2026-08-03)

• Gemini Robotics 2 entered — added Google DeepMind Gemini Robotics 2 (1.2T params, whole-body intelligent VLA, 22-DOF dexterous-hand manipulation, three-piece set: GR2+ER2+On-Device 2), robot_ai_models reached 21 entries
• Data scale expansion — chips 95→103, sensors 46→62, data acquisition devices +15 entries, total 493 entities covering 10 major categories
• July VLA model additions — added τ(0)-VLA (Shanghai Innovation Academy / AgiBot Robotics, slow-thinking fast-execution hierarchical architecture), InternVLA-A1.5 (Shanghai AI Lab, compositional generalization), Evo-Depth (SJTU MINT, lightweight 0.9B)
• WAIC 2026 model entries — added 5 new WAIC 2026 models: Hy-Embodied VLA-0.5, MiniCPM-Robot, Kairos 3.1, LingBot-VLA 2.0, Qwen-RobotManip
• Data categorization improved — added independent categories flexible_actuators (6 entries), robot_ai_models (21 entries) and data_acquisition (15 entries), total 10 major categories 493 entities
• Bionic mechanism category — added independent category bionic_mechanisms (9 entries), covering bionic joints, bionic actuators, bionic sensors, etc.
• VLA model updates — added GR00T N1.7, SmolVLA, π0.5, total LLM count reached 30
• Design canvas upgrade — integrated URDF Loader, supporting importing URDF files for real-time rendering and joint control
• Search engine upgrade — weighted fuzzy search, multi-field weight ranking + match highlighting
• Selection engine upgrade — five-dimension scoring system (torque / speed / accuracy / weight / cost) + multi-result comparison table
• Data pipeline automation — created data crawling and update automation scripts
• SEO automation — implemented automatic SEO metadata generation and management

8. API

• GET /api/entities.json — full entity list
• GET /api/compatibility_matrix.json — compatibility matrix
• GET /api/entity/{id} — single entity detail (free-tier fields + explicitly listed locked paid fields)
• GET /api/search?q=keyword — keyword search (optional category / limit / include_quarantine)

9. Deployment

• Platform: Cloudflare Pages
• Project name: robotparts
• Live URL（正式入口，外链一律指向此域）: https://roboparts.cc
• 线上地址（正式入口，外链一律指向此域）: https://roboparts.cc
• Preview domain / 预览域（Cloudflare Pages 默认域，非正式入口，勿对外引用）: https://robotparts-924.pages.dev

10. Security

• API Token has been removed from the docs; please manage it via environment variables
• The old token has leaked; please rotate it in the Cloudflare Dashboard

11. Quality Gates

The numbers in this repo and its externally published artifacts are computed live by scripts from the single source of truth api/entities.json; hand-editing is forbidden.

python scripts/ci_gate.py --list   # list all gates
python scripts/ci_gate.py          # run all gates (this is what GitHub Actions runs)

Total 78 gates:

Semantic index covers all entities · entity schema contract · mount_type enum contract · standard_conformance coverage consistency · external dataset distribution consistency · agent-discovery skill list consistency · MCP category coverage (stdio + hosted ↔ entities.json) · MCP package integrity · public list numbers computed live · kinematic reachability (yin-yang self-test + drift) · demand-signal discrimination layer (three-state fail-closed + external wording) · external JSON parseable · entities.json meta consistency · meta single source of truth · Functions top-level security · GitHub config YAML parseable · no credential leak · BOM assembly-order topological sort · feedback-signal backflow aggregation · flywheel idempotent / recoverable.

The gate count is injected live by scripts/inject_readme_stats.py reading the actual output line count of ci_gate.py --list. Previously hand-written as "8 items", the real value was already 20 — the 12 under-reported items were not "not done", they were simply not counted.

For the complete cross-file consistency regression (seven-number single-source, site-wide quantity assertions, external surface wording), see python scripts/regression.py.

After changing data, regenerate derived files: python scripts/regen_derived.py.

Cross-dimensional credibility signals

Beyond the gate count, three signals make this project verifiable beyond "we say it works":

• Mutation score — scripts/mutation_score.py generates 6 mutant variants (M1~M6) targeting known failure modes (three-state conservation break / axis marginal / evidence pointer / honest boundary / denominator shrink) and runs them against the existing 46 gates. Kill rate 1.00 (6/6 ≥ 0.85 threshold). Metric lives at api/mutation-score.json.
• Verification antipatterns — docs/verification-antipatterns-2026.md distills 12 incident-driven rules (AP-01~AP-12) from 170+ runs in ops/results/_SUMMARY.md. Each rule has a falsifiable negative test mapped to a verify_*.py self-test.
• Negative results — docs/negative-results.md records 10 explicitly-failed hypotheses (NR-01~NR-10) across three layers (evidence / distribution / metrics), each with a "re-route recommendation" so the failure is reusable rather than a dead end.

These three signals are wired into the site as public URLs and are surfaced in llms.txt's "工程可信信号" section so that LLMs discovering RoboParts via search see the verifiability story, not just the entity count.

12. Contributing

The biggest gap in this project is the mechanical interface declaration rate — for most cases we cannot answer "can these two parts screw together". This number is not hand-written here (see the computed value on the "Data volume" line at the top of this document); adding one piece of hole-pattern data with a source is far more useful than refactoring an algorithm.

• Add a mechanical interface declaration (no need to know JSON; just paste a source link)
• Report a data error
• Detailed rules and minimal verifiable format: CONTRIBUTING.md
• Authoritative format spec for part entities (field table + copyable template + gate list): docs/part-annotation-spec.md

The only hard rule: no source, no entry. We would rather keep not_declared than guess.

13. License

Dual-track licensing:

These two were previously written in the same LICENSE file; GitHub's license recognizer could not parse the mixed text ⇒ the repo page license field long showed NOASSERTION. They are now split: LICENSE holds only the verbatim MIT full text (recognizable), and the data line is moved to DATA-LICENSE.md.

Before citing the data, please note: the parameters in this repo are vendor public-stated values, not reproduced by our own testing; the number of A-grade entries that can be directly compared across vendors is 0. Each data point carries source_tier (A/B/C) and confidence; please judge credibility accordingly.

Generated: 2026-08-03 (sections 11–13 added 2026-08-31)