Business Plan · Embodied Robots

Business-planning and research materials for robotics (grouped by theme)

Download business-plan template (Word)
On this page

Embodied AI · 10-Person Team Business Plan

  1. Updated

    Existing Assets Summary: IVIS's Four Embodied-AI Demos and the Reservoir Computing Technology Foundation

    Before launching the venture, get clear on 'what we already hold.' This piece dissects, one by one, the four demos the company exhibited at NexTech Week 2026—① quadruped-robot contactless-sensing collision detection, ② Reservoir Computing for efficient control (rotary inverted pendulum), ③ edge-FPGA gesture recognition, and ④ edge speaker recognition—and distills the single technology foundation running through all four: Reservoir Computing (RC / ESN) as an 'ultra-lightweight, low-power, online-incremental-learning, time-series-anomaly-detection-savvy' edge AI. This capability layer (not 'buying a robot dog') is the true moat of this 10-person team, and the starting point for the later judgment of 'where to make money.'

    已有成果リザバー计算储备池计算Reservoir Computing边缘AIフィジカルAINEDONexTech
  2. Updated

    Embodied AI Direction Study: Starting From Our 'Reservoir Computing Moat,' Finding a 3-Year Direction That Both Earns Money and Survives

    The core question for launching this venture: a 10-person team, standing on the company's 'Reservoir Computing (RC) edge-AI' moat and starting from our 'existing results' — over the next 3 years, where should we go to both make money and stay alive? This piece first runs a bubble check on the Embodied AI / Physical AI mega-trend using external market intelligence (whole-machine/humanoid is a capital black hole — investment banks' TAM estimates differ by ~180x, Tesla itself admits Optimus is 'for learning, not production'; predictive maintenance / edge anomaly detection is the near-term cash, backed by an installed base of 4.66M industrial robots), adds a Japan demand 'confidence-grade' table (logistics/inspection/agriculture ★★★, eldercare ⚠️ overhyped), then aims the RC moat at the market layers (including two adjacent service layers — SI / functional-safety consulting — plus the Veo cash-burn cautionary tale), rapidly assesses and scores 5 candidate directions, squarely checks the competitive reality (TinyML/Augury/Senseye/TDK), and anchors Direction E to NEDO's AIRoA ¥20.5B foundation-model & data-platform program (2025–2029, overlapping our 3-year horizon). It closes with a 'main line + side bets' wager portfolio + a contract-work → productization → recurring 3-year roadmap + a 10-person staffing plan + KPIs + failure-mode avoidance. ⚠️ Market figures come from third-party research firms with widely divergent definitions; each is sourced inline, and the conclusions use only 'orders of magnitude,' never 'precise values.'

    方向性调研具身AIフィジカルAI预测性维护边缘AIリザバー计算NEDO商业模式3年路线
  3. Updated

    术语补充解释：RC / リザバー / エッジFPGA / リッジ回帰 / 世界模型 等

    配套 [已有成果总结](/robot-plan/team-assets-rc) 的『术语小词典』：把 IVIS 展示资料里出现、但不熟悉的关键词逐个讲清楚——RC（储备池计算）、リザバー、储备池/制御量的日语读音、エッジFPGA、リッジ回帰（岭回归）、世界模型学习。内容整理自向 Google 搜索 AI 模式提问得到的问答（已清掉图片与多余寒暄，保留解释与来源链接）。读 Demo 文档时遇到生词，来这里查。

    术语补充リザバー计算Reservoir ComputingESNエッジFPGAリッジ回帰岭回归世界模型制御量日语读音

Planning Methodology

  1. Updated

    [Feature] How to Write an Embodied-Robot Business Plan — Key Points, Intel Checklist, Fill-in Framework

    Written for an early-stage, exploratory venture: a software company + a partner company with factory ties + chairman/CEO as investors. The real purpose of this plan is not a polished execution roadmap but a 'go/no-go + investment-decision document.' This feature lays out the 7 core questions any review will ask, the 5 categories of intel to gather (with a ready-to-send request for a colleague), the persuasion logic chain, and an 11-section fill-in template (Japanese section names + plain-language guidance).

    专题事业企划具身机器人投资判断

Robot Dogs (Quadrupeds)

  1. Updated

    Robot Dogs (Quadrupeds): Vendor / Price / Performance Comparison + Selection Lens (Draft, To Verify)

    For the boss's idea — buy a low-cost quadruped chassis as a platform, mount functions on top, exhibit it at the Otsuka Shokai trade show, and use it day-to-day — here's a vendor comparison: split by price into budget/dev, industrial, and premium tiers, listing vendor, price, payload, runtime, ingress protection, and SDK openness, with a 'low-cost platform + add-ons' selection view. ⚠️ Prices are mostly USD reference values (some conflict across sources); Japan procurement channels and certification are unverified — please check and edit.

    机器狗四足机器人选型大冢商会平台化
  2. Updated

    AI & Robot-Dog Trade Shows (H2 2026 · Japan & China, for Partnership Talks)

    For the boss's goal of going to trade shows to talk partnerships with vendors, a roundup of AI / robotics (incl. quadruped) shows in Japan and China after 2026-06-25: name, dates, venue, highlights, official links. China-side WAIC / World Robot Conference / CIIF put you face-to-face with robot-dog makers like Unitree; Japan-side CEATEC / RoboDEX / Japan Robot Week suit reaching distributors, integrators, and customers. ⚠️ Some dates are unconfirmed (flagged) — defer to official sites.

    展会AI机器狗日本中国商务合作
  3. Updated

    Robot Dogs for In-Plant Haulage: Replace/Complement AGV·AMR — Payload × Stair-Climbing × Price Comparison (Draft, To Verify)

    Otsuka Shokai currently uses AGV/AMR to haul goods inside the plant, but they can't climb stairs/ramps; the boss wants robot dogs to cover that blind spot and pitch a partnership. The per-item weight of the goods is unknown so far, so this piece sorts every robot dog into a 'payload' tier and lays their walking payload · stair-climbing · speed · runtime · price into a fill-in-the-blank selection table to handle future weight changes. Conclusion first: robot dogs and AGV/AMR are 'complement, not replacement' — AGV/AMR carry big loads (100–1,500kg+) but only on flat floors, while robot dogs carry small loads (walking ≤~40kg) but climb stairs and ramps; the robot dog's unique value = the cross-floor/ramp/step 'last leg' of haulage. This edition adds a 'load + stair-climbing: is it actually stable?' section: vendor specs (continuous walking load vs standing peak load · step height · slope) + support-polygon/ZMP engineering basics, yielding a judgment method and an on-site test checklist. Unitree B2/B2-W figures have been cross-checked and filled in against Unitree's Chinese official site (unitree.com/cn) and overseas English official site. ⚠️ Prices are public-channel reference values, and Japan procurement/certification plus the real goods weight are to be verified; DEEP Robotics' site (deeprobotics.cn) had its certificate expire 2026-06-25 and is temporarily unreachable, so its figures are marked to-be-verified. Whether another type (tracked carrier/humanoid/goods lift, etc.) might fit this case better than a robot dog is covered in the sibling piece, Beyond the Robot Dog for Stair Haulage.

    机器狗厂内搬运AGVAMR载荷选型大冢商会
  4. Updated

    Quadruped × AGV/AMR Fusion: Products & R&D That Cure Both 'Legs Can't Haul' and 'Wheels Can't Climb Stairs' (working draft)

    Earlier articles framed robot dogs and AGV/AMR as 'complementary, not substitutes.' This one answers the boss's new question head-on: are there products or research that FUSE the two — or re-engineer a quadruped INTO an AGV/AMR — curing both pain points at once: quadruped low payload + AGV/AMR can't climb stairs/ramps? Yes — along three clear routes. (1) Wheeled-legged integration (wheels on the feet): roll like an AMR on flat ground, switch to a legged gait at stairs. This is today's mainstream answer — Unitree B2-W, Pudu D5, RIVR (formerly Swiss-Mile), Direct-Drive Tita, Tencent Max. (2) Carrier / car-dog cooperation (a heavy AMR ferries the quadruped on the trunk run; the dog dismounts for the stair-climbing 'last 100 m') — DeepRobotics Jueying X30/X20, GuoZi Robot, Unitree B2 and Boston Dynamics Spot all do this, but mostly as integrator/project-level solutions with no off-the-shelf 'car-dog' SKU. (3) Heavy-payload HYDRAULIC quadrupeds (SJTU 'Baby Elephant', NUDT) that push pure-leg payload past 100 kg. We sort all of it into 'buy-it-now / R&D prototype / concept-patent' tiers and check differentiation & limits (payload, terrain, cost, maturity, and whether the dispatch stack aligns with the AGV standard VDA5050). Key Japan lead: Unitree B2-W has an official distributor, TechShare (techshare.co.jp) — a real procurement channel for the Otsuka case. ⚠️ This merges our deep-research workflow (verification partly cut off by a session limit) with the user's Google-AI-mode research doc; two of the doc's claims are corrected (the 'Unitree W1' is really the B2-W; 'Tita' was wrongly attributed to Minimotors) — see the errata. Most specs are vendor/press figures; verify with on-site loaded testing before committing.

    机器狗AGVAMR轮足融合重载四足子母车大冢商会
  5. Updated

    Beyond the Robot Dog for Stair Haulage: Tracked Carriers · Humanoids · Stair-Climbing AMRs · Goods Lifts — A Horizontal Review (Draft, To Verify)

    Otsuka Shokai's pain point is 'goods need to cross floors / climb ramps, but AGV/AMR can't manage stairs.' Robot dogs can cover that blind spot, but their walking payload is only ~40kg. So is there a robot better suited to this than a robot dog? This piece puts five categories side by side — wheeled-legged platforms · tracked electric stair-climbing carts · humanoids · stair-climbing AMRs · non-robot solutions (goods lifts/VRC/conveyors). Honest conclusions: ① a truly 'autonomous + climbs real stairs + carries load' mature product barely exists — what you can buy is essentially a 'robot dog on wheels' (B2-W, Pudu D5); ② everything that crushes the robot dog on payload (200–420kg) is a tracked/electric stair-climbing cart, but all need a human operator; ③ as of mid-2026 no humanoid can carry loads up stairs on a production line; ④ for a fixed inter-floor route with real stairs, the most rational answer is often not a robot at all, but a goods lift/VRC or 'AMR on the flat + elevator linkage' — cheaper and more reliable. Robot dogs still win on rugged/narrow/multi-purpose/low-load variable-route work. ⚠️ Most specs are vendor claims or media relays; before deployment verify step height and slope under load.

    机器狗爬楼搬运履带车人形机器人AMR货梯VRC选型大冢商会
  6. Updated

    Mother-Child Vehicle (子母车) Research Review + Otsuka Solution: marsupial heterogeneous robot cooperation — cases, papers, results (working draft)

    The boss wants the existing cases, concept papers, and research on '子母车 (mother-child vehicle),' then a direction & solution for Otsuka's real need. First, the name: '子母车 / car-dog cooperation' is academically marsupial robotics (a larger carrier/mother robot carries and releases a smaller passenger/daughter) — a mature sub-field of heterogeneous multi-robot cooperation. This article gathers three blocks: (1) concepts & surveys (the classic marsupial-definition survey + the open-access ACM survey of cooperative heterogeneous multi-robot systems); (2) the most on-topic results (quadruped-carrier + child deployment criteria arXiv 2205.05477; deadlock-free scheduling of attachable heterogeneous carrier-shuttle AGVs arXiv 2508.00724 = the industrial 子母车; ETH's 85 kg legged-suspension carrier LEVA; dynamic docking on complex terrain; tethered marsupial power for endurance); (3) industrial deployments — the honest finding: the ONLY publicly verifiable 'autonomous vehicle carrying a robot dog' deployment worldwide is one (UISEE × State Grid Hangzhou 'car-dog integrated,' 2025, where an L4 vehicle ferries cross-zone and the dog dismounts to climb stairs / enter utility tunnels for last-100-m inspection), while Spot/ANYmal/DeepRobotics cases are all 'solo quadruped + charging dock,' NOT car-dog (WebFetch-verified, don't misuse); the broad 子母车 has mature four-way shuttle-carrier ASRS (Dematic, Inform). Finally it maps the research to Otsuka and proposes a 'dual-track (wheeled-legged solo / carrier cooperation) + phased PoC + VDA5050 software-orchestration differentiation' direction. ⚠️ All links curl-verified reachable; arXiv 2202.08620 (Kirin) was withdrawn by its authors and is NOT Baby Elephant — not used as a credible result. Publisher pages (IEEE/ScienceDirect) returning curl 403 are bot-blocks, not dead links, so arXiv/open PDFs are cited preferentially.

    机器狗子母车marsupial异构协同AGVAMR车犬协同大冢商会方案
  7. Updated

    Starting with Unitree GO2: GO2 lineup payload × price (domestic + overseas) × carrier/AMR integration reality and a starter plan (working draft)

    The boss wants to start planning with the Unitree GO2. This article focuses on GO2 only: the payload & obstacle-crossing of the whole lineup (Air/Pro/EDU + wheeled GO2-W), prices on the domestic official site + overseas channels, and the carrier-vehicle question — which AMRs/AGVs has the GO2 actually been integrated with. Three honest conclusions up front: (1) GO2 is consumer/education tier, haulage capacity ~7–8 kg (regular/rated); the official 'peak ~10–12 kg' is a peak limit, NOT haulage capacity — don't promise on it; the official spec lists NO IP rating and NO 'loaded vs unloaded' step/slope figures, so all step/slope here are unloaded nominal. (2) Price: domestically GO2 Air from ~9,997 RMB, Pro ~18,000 RMB, EDU on quote; overseas shop.unitree.com from ~$2,800, EDU on quote (specs per www.unitree.com; shop has price only, no spec table). (3) Carrier-vehicle, the key one: as of 2026-06 the GO2 has NO ready-made AMR/AGV system integration or car-dog deployment — you must self-build middleware on the official unitree_sdk2 / unitree_ros2 / DDS; GO2 does NOT support VDA5050; and only the EDU version opens full secondary development (incl. low-level motion control) — Air/Pro consumer versions are essentially locked. Conclusion: starting with GO2 means EDU only, positioned as a low-cost PoC of '≤7 kg light load + stair-climbing + in-house dispatch software,' with integration self-built — which is exactly the in-house-software differentiation entry point. ⚠️ All links curl-verified reachable; RMB prices are media/e-commerce figures that fluctuate — re-check before purchase.

    机器狗宇树GO2负载价格子母车AMRAGVSDK大冢商会方案
  8. Updated

    Robot dog × AMR/AGV system integration & car-dog cooperation: who's actually wired up — a global case list (working draft)

    Answers one specific question: which concrete AMRs/AGVs have robot dogs actually been system-integrated / interface-bridged / partnered with? The honest picture after curl + WebFetch verification — (1) real 'mobile carrier ferries/releases a child dog' commercial integrations are very few worldwide: in China only UISEE autonomous vehicle × State Grid Hangzhou (vehicle ferries dog cross-zone, dog does the last 100 m), overseas only RIVR (ex-Swiss-Mile) wheeled-legged × delivery vans (Evri UK / Veho US, dog goes from van all the way to the doorstep). (2) The most solid 'quadruped plugged into an AMR dispatch system (RCS)' is SEER (Standard Robots): robot dogs D1/D2 plugged into the M4 dispatcher, SRC controllers driving third-party quadrupeds (note: NOT VDA5050). (3) Inter-model cooperation (not car-dog but related): DeepRobotics Jueying X30 + Lynx M20 wheeled-legged, GuoZi/Leaderobot multi-type (wheeled/rail/quadruped/drone) into one inspection host. (4) Common misreads to clear up: Boston Dynamics Spot Dock, Asylon DogHouse, ANYbotics ANYmal dock are all 'quadruped walks back to a FIXED charging dock,' not carrier cooperation; wheeled-legged machines (B2-W/Lynx) have their own leg-end wheels, not an external AMR. (5) Negative matrix: Unitree has NO named third-party AMR integration; Hikrobot/Geek+/YOUIBOT/Standard/Forwardx/Siasun (6 vendors) are all AMR+arm with ZERO quadruped integration; China's VDA5050 deployments are all wheeled AMRs, no robot dog; openTCS/OTTO fleet managers don't yet include quadrupeds. Implication for Otsuka: robot-dog×AMR integration is almost entirely self-built; SEER M4 is a domestic reference path for 'quadruped into dispatch'; don't expect off-the-shelf VDA5050 — which is exactly the in-house-software differentiation space. ⚠️ Links curl-verified; some news sites (FreightWaves/RobotReport) need a browser UA, bare curl returns 403, not a dead link; mainland gov sites returning curl 000 are geo-blocks, not dead links.

    机器狗子母车车犬协同AMRAGV系统集成RCSVDA5050仙工SEER驭势RIVR大冢商会
  9. Updated

    Marsupial-Shuttle Warehouse Transport on SEER's System — A Feasibility Study (downloads-center first-hand check × Japanese support × SRC heterogeneous compatibility × GO2/B2 fit · working draft)

    Following the SEER deep-dive: pinning down whether 'marsupial-shuttle warehouse transport on SEER's system' can actually land. Four points, each first-hand checked — (1) SEER's downloads center (seer-robotics.ai/zh/download, curl 200) really hosts product catalogues/videos, and there is a dedicated Japanese site + Japanese downloads center (/jp/download); (2) Japanese support is a natural plus for Otsuka: the corporate site offers 9 languages incl. 日本語 and a JP site — but 'the M4 product UI is localized to Japanese' could NOT be directly verified, and SEER has NO local agent in Japan (inquiries route to HQ), so soften the wording; (3) retrofitting SEER's SRC controller to make heterogeneous dogs/AMRs compatible is a real path: a third-party dog/AMR with an SRC inside plugs seamlessly into M4 unified dispatch (official '1,000+ customers' / media '1,300+ partners'); for legged/embodied use ONLY SRC-5000 applies (note: SRC-1100 is a dual-steering-wheel controller, NOT for quadrupeds); (4) Unitree GO2 vs B2 fit with M4: GO2 opens SDK only on EDU, ~7 kg light load; B2 fully opens SDK, walking load >40 kg (the 120 kg standing figure is static, not haulage) — both need self-built middleware or a bolted-on SRC-5000. Marsupial form = M4 as the unified RCS: mother car (AGV/AMR via SRC-native or VDA5050) + child dog (SRC-5000 brain) co-dispatched; download the manuals first to verify first-hand, then PoC. Three honest gates kept: SEER has no local agent in Japan; VDA5050 depends on Otsuka's existing AGVs (and the standard only targets 2D wheeled AMRs); GO2 light load, B2 heavy load. ⚠️ Errata: SEER's D1/D2 are 'partner hardware + SEER brain,' NOT SEER-built units, media-named; Unitree's D1 is a 6-DOF robotic ARM, not a dog — keep the two apart. All links curl-verified 200.

    仙工SEERM4SRC子母车仓储运输VDA5050日语宇树GO2B2大冢商会方案
  10. Updated

    Deconstructing the SEER System-Software Framework — A Reference Benchmark for Our In-House «Dog + AGV/AMR Coordination System Software» (Layered Architecture × External Interfaces × What to Borrow vs How Our In-House Build Maps to It · Working Draft Under Continuous Verification)

    Strategy has pivoted: the coordination system software for the robot dog and the AGV/AMR fleet will be built in-house by the company; SEER is not a procurement target but a reference benchmark — this piece digests its system-software framework to serve as a reference for our in-house build. After verifying each official page, we tease out a five-layer architecture: L1 sensing/actuation → L2 per-robot autonomy (SRC controller: mapping/localization/obstacle-avoidance/motion-control pushed down into the controller) → L3 the dispatch brain (M4 RCS: dynamic order assignment/routing/traffic control with collision-detection + Safe Swap; stack = JVM + Rust/C++, React frontend, dual extension via Python backend scripts / JS frontend scripts) → L4 business integration (fully open HTTP API + WebSocket + callback to connect ERP; Modbus/OPC/S7/VDA5050 are solution-level claims to be verified) → L5 toolchain (Roboshop Pro low-code per-robot implementation / Meta-World digital twin / Nebula online configuration). The core design philosophy = «push per-robot autonomy down into the SRC, lift fleet dispatch up into the M4, decouple the two layers» — this is the single most worthwhile thing for our in-house build to copy. Among the external interfaces, the five items API/WebSocket/callback/Python/JS are explicitly stated on official product pages and are trustworthy; the concrete Modbus/OPC/S7/VDA5050 lists are not fabricated and are marked to be verified. We provide a «SEER design point → how our in-house build maps to it» comparison table, and call out four limitations that are exactly our in-house opportunity windows: it targets only 2D wheeled vehicles (a dog's legged obstacle-crossing / 3D terrain is not in its dispatch model), VDA5050 is not stated on the M4 homepage and only covers planar vehicles, the SEER D1/D2 quadrupeds are not self-built complete machines (partner hardware + SEER brain SRC-5000), and there is no local agent in Japan. Errata locked in: the only controller for a dog is the SRC-5000 (the SRC-1100 is a dual-steering-wheel controller, not for quadrupeds); the Unitree D1 is a six-axis robotic arm, not a dog.

    仙工SEERM4SRCRoboshopMeta-WorldNebulaVDA5050系统架构对外接口自研大冢商会参考标杆
  11. Updated

    Robot Dog ↔ AGV/AMR Coordinated Dispatch System Software · Comparative Reference List — Who Else Besides SEER Is Worth Studying (Open-Source Frameworks openTCS / Open-RMF × VDA5050 Standard × Commercial Benchmarks × Domestic RCS × Single-Robot Navigation Nav2 · All External Links curl-Verified)

    Strategy has shifted toward developing our own coordination system software for robot dogs and AGV/AMR, with SEER being just one reference benchmark. This piece sweeps across dispatch/coordination systems worth studying at home and abroad, giving each one an entry on 'architecture / external interfaces / open-source or not / can it be borrowed / can quadrupeds be included.' All external links are curl-tested. [The two open-source frameworks most worth copying] ① openTCS (Fraunhofer IML, MIT): the Kernel + Plant Overview + Kernel Control Center trio, with pluggable Dispatcher/Router/Scheduler strategies, one communication adapter per vehicle type + loopback simulation, external RMI + HTTP/JSON Web API (OpenAPI + API Key) + event stream, ships an official VDA5050 adapter; ② Open-RMF (Open Robotics, Apache-2.0, built on ROS 2): the most direct template for heterogeneous multi-vendor fleet interoperability — Fleet Adapter + three-tier integration (Full Control/Traffic Light/Read-Only/Mixed), centralized traffic management rmf_traffic, task allocation rmf_task, battery modeling rmf_battery, infrastructure abstraction for doors/lifts/dispensers; robot dogs can join a shared pool with wheeled AMRs via a custom adapter = the key advantage over openTCS/VDA5050. [Communication standard] VDA5050 (v3.0.0, MQTT+JSON, six topics, node/edge graph + base/horizon) targets only 2D wheeled robots; open-source implementations libVDA5050++ / InOrbit connector / berketunckal master-control reference are worth borrowing — quadrupeds should borrow the semantic skeleton rather than apply it directly. [Commercial benchmarks] OTTO (Rockwell, closed-source Fleet Manager + open-source VDA5050 connector), MiR (under Teradyne, REST API + open-source VDA5050 adapter), Spot Orbit (REST+Webhook HMAC-SHA256+Bearer = the gold template for external interfaces), ANYmal (RESTful, productized inspection data), DEEP Robotics/Unitree (only single-robot SDKs, no off-the-shelf fleet software = confirming the self-development gap). [Domestic RCS] HAI Robotics HAIQ / Quicktron Quick / Geek+ RMS (multi-form same-pool · closest to heterogeneous) / ForwardX (markerless vision navigation) / Hikrobot (led national standard GB/T 43047-2023) / Enotek are all closed-source, only dispatch their own vehicles, and don't support quadrupeds. [Single-robot layer] Nav2 (Apache-2.0, ROS 2 Action, officially supports legged chassis) — use it directly, don't reinvent. Overall recommendation for self-development: business (WMS/MES) → custom coordination kernel (RCS/FMS task dispatch + traffic control + avoidance + charging) → Fleet Adapter (one per machine class) → single-robot Nav2 → vendor gait stack; copy Open-RMF's Fleet Adapter three-tier integration + openTCS's strategy abstraction/communication adapter/loopback simulation + Orbit's REST+Webhook external interface; southbound to AGVs via VDA5050, to dogs via a custom adapter (extending gait/posture/obstacle-crossing fields).

    openTCSOpen-RMFVDA5050Nav2RCSFMSOTTOMiRSpot OrbitANYmal极智嘉海柔自研大冢商会横向参考
  12. Updated

    Marsupial (Mother-Child) Vehicle Collaboration: After Building Our Own System Software, Do We Still Need Add-On Hardware Between the Child/Mother Vehicles?——Software Interface vs. Physical Docking Boundary × Which AGV/AMR Unitree GO2/B2 Can Dock With Add-On-Free × Cost Order-of-Magnitude (Working Draft to Be Verified)

    The boss's core question: after the company builds its own [marsupial-vehicle collaboration system software], do the child and mother vehicles still need add-on hardware between them? Or is defining the interface enough to skip it? One-line answer——[it depends on whether you want the mother vehicle to 'physically carry' the child dog]. First, clear up a misconception: a software interface defines the 'rules of conversation' (API/protocol/scheduling timing), while physical docking does the 'carrying actions' (carrying/alignment/locking/charging); software can't move atoms—it can only command hardware and read hardware feedback, so 'defining the interface' ≠ 'no hardware needed.' We break hardware into five categories and go through each: ① communication hardware, ② edge-compute boxes—these two are built into Unitree GO2 EDU/B2 (gigabit Ethernet port/Wi-Fi/Orin·i7 compute), so as a child vehicle it basically needs no add-ons and connects directly via the software interface; ③ localization/alignment sensors—the dog's own LiDAR/cameras can be reused, but 'precise mother-vehicle docking + physical dog-vehicle mating' still often needs dedicated alignment markers (AprilTag) and presence detection; ④ mechanical carrier deck/ramp/locking mechanism/charging contacts—unavoidable as long as you 'physically carry,' and software can't replace them; ⑤ safety hardware-cutoff I/O (safety bumper/light curtain/e-stop)—strongly recommended, and it dovetails neatly with the company's RC edge anomaly-detection moat. This gives two routes: Plan A true marsupial vehicle (mother physically carries the child dog) = mechanical docking hardware is unavoidable; Plan B logical relay (the dog walks on its own, the AGV drives on its own, software only schedules the handoff) = even mechanical docking is saved, best for a software company to start with. The mother vehicles Unitree GO2/B2 can directly connect to as a child vehicle: ROS2-native chassis (AgileX/Clearpath, one Ethernet cable, ROS2 Topic straight through), commercial AMRs with open HTTP WebAPI/Modbus TCP (MiR/Geek+/Hikrobot/Quicktron, send JSON commands), and I/O GPIO hard-coupling (B2 has multiple 24V/12V/GPIO lines, physical contacts give the mother vehicle a 0-latency high/low logic level for safety lockout). Costs are order-of-magnitude estimates only (communication ¥3k–15k / LiDAR ¥8k–40k / edge FPGA box ¥2k–10k / safety mechanism ¥1.5k–6k). Honest corrections: only Unitree's EDU edition opens full low-level secondary development (consumer Air/Pro are basically locked); the exact interface counts and compute of GO2/B2 should follow the official spec pages, and prices vary by channel and time; VDA5050 only solves logical-layer interoperability and handles no mechanical docking whatsoever.

    子母车车犬协同外接硬件软件接口VDA5050ROS2宇树GO2B2AGVAMR边缘异常检测NEDO大冢商会自研

Capability Research

  1. Updated

    Robot-Dog Capability Map: 11 Capability Axes × Official Quotes × Task Matching (Draft, To Verify)

    The boss needs to judge "what's worth building" — not just list capabilities. This piece places robot dogs on a unified 11-axis capability frame (mobility / payload / manipulation dexterity / perception / autonomy / human-robot interaction / runtime / dev openness / reliability maturity / cost / certification) to paint a strengths-and-weaknesses profile, backs each conclusion with a verbatim quote from the vendor's official page (with a clickable footnote superscript at the end jumping to the source), and lands on the judgment of "which tasks these capabilities suit / don't suit." Conclusion first: a robot dog = an excellent autonomous mobile sensing platform — strong in mobility / obstacle-crossing / payload / runtime / low cost / mature outdoor inspection / SDK ecosystem, weak in dexterous manipulation (unless an arm is added) and human-facing interaction — inspection, transport, and data collection are its home turf. ⚠️ All quotes are verbatim from official pages (sourced line-by-line), but price / procurement / certification are unverified and a few specs aren't on the official page (flagged). Please verify and edit.

    机器狗四足能力轴具身AI选型任务匹配
  2. Updated

    Other Embodied-AI Atlas: 6 Form Factors × 11 Capability Axes × Task Matching (Draft, To Verify)

    Robot dogs are just one form factor of embodied AI (Physical AI). This piece puts six form factors — humanoid / wheeled service / robot arm / dexterous hand / wheeled-legged / drone — onto the very same 11 capability axes used for robot dogs: first a '6 forms × 11 axes' horizontal portrait matrix, then each form factor backed by official verbatim quotes + clickable footnotes, finally landing on task → most cost-effective form factor: inspection/hauling → quadruped, need human tools + two hands → humanoid, pure interaction → wheeled with screen, fixed-station precision → robot arm, aerial → drone. ⚠️ All quotes are official text (sourced line-by-line); Tesla's whole domain returns 403 to the fetcher, so related content is flagged UNVERIFIED; the direct robot-dog-vs-humanoid comparison has no official source and is flagged 'analysis (to verify).'

    具身AIフィジカルAI人形灵巧手机械臂无人机形态对比任务匹配
  3. Updated

    Candidate Entry Scenarios · Feasibility Quick-Read: Running the Numbers on 3 Scenarios + Priority (Draft, To Verify)

    The first two pieces nailed down the capabilities; this one helps the boss decide 'is it worth doing.' Three candidate scenarios are picked — (a) facility inspection = robot dog + in-house inspection software, (b) expo/reception interaction robot, (c) mobile data-collection platform — each rated on the same 5 items: ① which form factor (per the form-factor-fit conclusions) ② rough cost (base + add-ons, drawing prices from the capability research, flagged to verify) ③ the unique value our own software/AI can add ④ Otsuka Shokai channel/customer fit ⑤ ramp-up difficulty and risk. It closes with a priority ranking + recommended first pick. Conclusion first: Scenario A (inspection) lands in the quadruped's all-green zone, has the highest in-house software value, and the smoothest channel — recommended as the first entry; C (data collection) as a same-base second step; B (interaction) needs a wheeled-with-screen form, has high AI value but strays from the robot-dog home turf, listed as an opportunistic pilot. ⚠️ Costs and channel fit are estimates to verify; please have the boss correct them with real quotes/customer data.

    切入场景可行性巡检交互机器人数据采集选型任务匹配
  4. Updated

    Where Robot Services Have Landed × Software Worth Borrowing/Integrating — 6 Deployed Domains × 5-Layer Software Stack (Open vs Closed) × Where We Can Build (Heterogeneous Fleet Middleware / Marsupial Coordination / Data & Eval Tools, all links verified)

    Answers two questions: ① which domains have actually deployed robot services, and ② what software those services use — and where a small software/integration team can break in. [Six landed domains] Auto/3C factories (AgiBot G2 on Longcheer's line: 8h, ≥99.5% success, 310 units/h; UBTECH Walker S at Zeekr/Dongfeng/BYD; Figure 02 @ BMW; Atlas @ Hyundai; Apollo @ Mercedes; Digit @ Schaeffler RaaS $30/h), logistics (Amazon fleet 1M+ units; Geek+ RMS; JD sortation 90% de-manned), commercial service (Keenon dining #1 at 22.7%; Pudu 130k units/80 countries; Gausium cleaning #1 IDC 12.9%; YunJi hotels 34k venues), inspection/security (Spot+Orbit AI vision; Unitree/DEEP Robotics substations), home companionship/after-sales (UBTECH U1; JD JoyRobocare EU repair network), and robot schools/data factories (Apptronik Robot Park; AgiBot 4000㎡ data factory; Shanghai national center; Beijing Yizhuang 1000-robot collection). [5-layer stack, open vs closed] Brain/VLA — free-to-use open: π0 (Apache, runs on a 4090), NVIDIA GR00T, LeRobot/SmolVLA; China's AgiBot GO-1, Galaxea G0, X Square WALL-OSS (all CC BY-NC, non-commercial); closed: Tesla, Figure Helix, Skild, Gemini Robotics, Noematrix, Galbot GroceryVLA. Sim = Isaac/MuJoCo/Genesis/Cosmos (commercial-OK). Data = ALOHA/UMI open hardware, Scale AI/Encord, JD's data exchange. Fleet middleware = VDA5050/Open-RMF/openTCS/MassRobotics, SEER RDS, Hikrobot RCS, InOrbit, Viam, Intrinsic. Service clouds = Keenon (KONE elevator API), Pudu PuduOS, Gausium FieldBots (already wired to Open-RMF + InOrbit), Agility Arc, Spot Orbit — all closed ecosystems. [Where we build] 🔴 avoid = robot bodies, training a base model from scratch, single-vendor FMS (Rocos/Freedom/Ready Robotics have exited); 🟢 worth it (by fit) = ① multi-vendor heterogeneous fleet orchestration middleware (incl. marsupial/shuttle coordination — no common protocol, a gap; continues our existing SEER work) ② data curation/QC/eval SaaS (force/tactile-vision alignment is the gap) ③ VLA eval-as-a-service + edge-inference middleware (a vLLM-for-robots) ④ vertical last-mile integration. Core logic = harvest the 'fragmented ecosystem + unsettled standards' dividend without touching hardware/compute/data moats. Key proof point = Gausium voluntarily wired itself into Open-RMF and InOrbit, showing heterogeneous integration is real demand. Starting point: 7 ifeng (Phoenix) reports.

    具身AIフィジカルAI落地场景VLAIsaac GR00Tπ0VDA5050Open-RMF异构调度子母车RaaS数据工厂切入点
  5. Updated

    Japan's Demand for Already-Landed Embodied-AI Tech × Where Our Software Fits — Otsuka Shokai channel needs + macro direction (labor shortage · Labor-Saving Investment Subsidy · RFA standards) + concrete pain points (warehouse utilization stuck at 60–65% · zero cross-brand fleet management in food service · siloed inspection data · care not wired to nurse call) + one match table

    Building on the previous piece «Landed Domains × Borrowable Software», this one investigates specifically [Japan]'s demand for already-landed tech, landing it in three blocks + one match table. [In one line] Japan's hardware (serving/cleaning/transport/inspection) is already amply supplied and deployed at scale; the entire gap is in the upper software layer — multi-brand unified orchestration + integration with existing IT · elevators · POS · nurse call · MES + O&M SaaS — which is precisely our chosen main direction ① (heterogeneous fleet orchestration middleware / marsupial coordination). [① Otsuka Shokai (focus)] Not a robot specialist reseller but Japan's largest independent IT SI + office trading house (~295k SMB clients · FY2025 consolidated revenue ¥1.3228 trillion), treating robots as a commodity it can 'bundle into a monthly subscription + O&M support'; already reselling temi / Pudu Mars / Preferred kachaka / Gausium Phantas / Makita / WinActor; explicitly wants multi-brand unified management + cloud single-pane management (temi Center) + connection to existing IT/ERP (SMILE V) + one-stop O&M (たよれーる / Tayoreru) — almost entirely overlapping entry point ①; the correct posture = give it a neutral multi-brand management middleware that can be white-labeled into Tayoreru. [② Macro direction] Structural labor shortage is irreversible (working-age population 2070→45.35M · shortfall of 11M by 2040 · labor-shortage bankruptcies hit a record high for two straight years, 350 cases); the government uses the Labor-Saving Investment Subsidy (new in 2024, ~¥500B · catalog 1/2 subsidy · catalog includes cleaning/serving/AGV·AMR) to directly subsidize SMBs buying robots; RFA has made elevator linkage + multi-robot fleet management into national standards (RFA B0001:2025) — policy is underwriting heterogeneous orchestration + systems integration. [③ Concrete pain points] Every industry shows the same pattern: single machines are mature → the moment you mix brands or connect systems, there's no neutral layer. Warehouse utilization stuck at 60–65%, zero cross-brand fleet management in food service (POS/KDS via custom builds), inspection data siloed per vendor cloud requiring third-party bridging, factories where Panasonic Robo Sync achieved multi-vendor single-pane control but doesn't connect to MES, care sensor aggregation platforms covering only the watch-over (見守り) category. [Match table] 7 ★★★ hits (warehouse/food service/factory/care/policy/Otsuka). [Honest caveat] The middleware space already has PLiBOT (Obayashi affiliate) / Robo Sync / Yokogawa RMC / VDA5050 staking claims; differentiate by focusing on safe, compliant integration of multi-vendor 【AMRs + elevators/POS/nurse call + unified O&M + white-labelable to SI channels】. Key external links have all been batch-curled, all reachable HTTP 200.

    具身AI日本市场大塚商会人手不足省力化投資補助金RFAVDA5050异构调度子母车介护物流2024問題切入点

AI · Robotics News

  1. Updated

    Embodied AI's “Skill Moment”: NVIDIA Open-Sources the ASPIRE Robot Skill Library, Jim Fan Says the Paradigm Has Changed

    NVIDIA GEAR open-sources ASPIRE, a robot skill library: like a coding agent, it distills every successful debugging fix into a reusable, retrievable skill; two-arm handover success climbs from 20% to 92%. Jim Fan calls it a paradigm shift — from training weights by gradient descent to an ever-growing skill library.

    具身智能英伟达ASPIRECode as PolicyJim Fan
  2. Updated

    A Looped World-Model Paper Tops Hugging Face: From Chinese Startup FaceMind, Backed by Zhou Hongyi and Lu Qi

    The Looped World Model (LoopWM) tops Hugging Face Papers: a parameter-shared Transformer iteratively refines latent state, for up to 100× parameter efficiency. Behind it is FaceMind Research Asia, a Chinese startup founded by two post-1995 PhDs, which closed a tens-of-millions-RMB Pre-A round led by Xinglian Capital, with 360 over-subscribing and Lu Qi's MiraclePlus participating.

    世界模型LoopWM脸谱心智融资陆奇周鸿祎