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

⚠️ A working draft under continuous verification. This merges two inputs: (1) our deep-research workflow (multi-source search + adversarial verification — but the verify/synthesize stage hit a session limit, so some claims have a source but no completed vote; those are tagged "to verify"); and (2) the user's real-time research doc from Google's browser AI mode (good currency, but still second-hand, and we've corrected two of its claims — see the errata). Items with a [source] badge cite a primary/press source we actually fetched; details on the Chinese makers (Tencent, Direct-Drive, SJTU, NUDT) come largely from the user's doc and second-hand reports — trust the official datasheet + on-site loaded testing before deployment.

🌐 What this answers (the boss's new question)

Earlier the conclusion was that robot dogs and AGV/AMR are complementary, not substitutes (see Haulage selection, Stair-climbing alternatives). This time the boss asks one step further:

Are there products or research that FUSE quadrupeds with AGV/AMR — or re-engineer a quadruped INTO an AGV/AMR? That is — curing both pain points at once:

• Quadrupeds have low payload (pure-leg walking payload generally ≤ ~40 kg);
• AGV/AMR can't climb stairs, ramps, or steps.

Yes — and it's one of the hottest directions in embodied AI + industrial logistics these two years. The industry calls this class of fusion product "wheeled-legged robots," plus two less-visible routes solving the same contradiction. Below: three fusion forms, then a three-tier maturity map for deployment.

🧭 Bottom line first

1. The mainstream answer = wheeled-legged integration (wheels on the feet). On flat ground, lock the legs and roll on the foot-wheels like an AMR (load becomes static support, motors no longer burn torque to stand+step → payload and range jump several-fold); at stairs/steps, lock the wheels and instantly switch to a legged gait. Buy-it-now examples: Unitree B2-W, Pudu D5, RIVR (formerly Swiss-Mile), Direct-Drive Tita.
2. Payload is still the ceiling. Wheeled-legged lifts "how much you can pull on flat ground" (B2-W tows ~100 kg class), but the walking payload you can carry UP stairs is still ~40 kg — wheels-on-legs ≠ carrying more up stairs (see the "walking vs standing-peak load" section in Haulage selection).
3. For true heavy payload (100 kg+), only two routes: (2) carrier/piggyback (a heavy AMR ferries the dog on the trunk run; the dog dismounts to climb stairs for the "last 100 m"), or (3) heavy hydraulic quadrupeds (SJTU "Baby Elephant," NUDT) — the latter mostly research/military prototypes, not commercial.
4. The software stack isn't aligned with AGV standards yet. These fusion machines almost all expose SDK / ROS2 interfaces and rarely claim VDA5050 compliance (the AGV fleet-master interop standard) — slotting into Otsuka's existing AGV master controller needs integration work (to verify; see limits).
5. Procurable in Japan: Unitree B2-W has an official distributor, TechShare (techshare.co.jp)¹¹— the most realistic procurement lead for the Otsuka case.

🧩 Three fusion forms (the mechanism that cures the pain)

Form 1: Wheeled-legged integration (wheels on the feet) — the mainstream endgame ✅ in production

Replace the quadruped's feet with actively driven wheels — one machine, two modes:

• Flat = AGV/AMR mode: legs locked, rolling on foot-wheels. Joints sit in static support, motors stop burning torque on "stand + step," so effective payload and range are several-fold a pure-leg gait, and it's faster (B2-W wheel speed reaches the 15–20 km/h class).
• Stairs/steps/steep ramps = legged mode: wheels lock, instantly switch to an RL-controlled legged gait and climb step by step — exactly what an ordinary AMR can never do.

Representative products (buy-it-now):

Why wheeled-legged is "the mainstream answer": it's the only form where one machine gets both "high payload + speed on flat ground (AMR strength)" and "climbing real stairs (legged strength)" — no vehicle swap, no human intervention. Our Stair-climbing alternatives reached the same verdict: what you can buy today that truly hauls up stairs is, in essence, "a robot dog with wheels."

Research frontier (same form, not commercial):

• Tencent Robotics X "Max" / "Ollie": Tencent's in-house wheeled-legged hybrids — roll fast and heavy on flat ground, hop or switch to a legged gait at double steps. Lab prototypes, not on sale (latest progress to verify).
• LEVA (ETH Zürich): ETH's RSL wheeled-legged logistics vehicle — parallel-kinematics legged suspension + steerable wheels, 85 kg payload, climbs stairs with an RL controller, and autonomously loads a cargo box under its body — putting "autonomous loading + obstacle-crossing transport" on one platform⁹. ICRA 2025 paper, research prototype, not a product. It shows the wheeled-legged payload ceiling is being pushed up.

Form 2: Carrier / car-dog cooperation (an AMR ferries the quadruped) — for true heavy payload, vendors already do this ✅🔬 mostly integration

When the load is hundreds of kg to tonnes and a wheeled-legged machine's "thin legs" can't bear it, the realistic answer is mother-child cooperation — also called "wheel-leg cooperation" or "car-dog / vehicle-dog cooperation" (车犬协同) in industry/academia. It's the hardware-integrated version of the "AGV runs flats + dog relays stairs/ramps" narrative in Haulage selection: the carrier provides "muscle and range," the quadruped the "last-100-m vertical crossing."

Four-stage workflow (the standard cycle for cross-floor intralogistics/inspection):

1. Heavy trunk run (carrier stage): the carrier (AGV/AMR) ferries the quadruped + cargo/inspection gear fast along flat main aisles; the dog sleeps/charges and draws no power of its own (curing the dog's short range, high power draw, light load).
2. Terrain switch (docking stage): the carrier drives to the "helipad" at the staircase / steep ramp / narrow passage.
3. Detached execution (dog stage): the dog wakes, walks off the carrier on its own, shoulders light sensors or small high-value goods, and climbs stairs / crosses thresholds into the 2nd floor or cross-floor zone for inspection / meter-reading / last-leg delivery.
4. Return & recharge (recovery stage): the dog returns to the helipad, climbs back onto the carrier and locks in, and the carrier ferries it to the next cycle.

Who does it (car-dog / composite-inspection solution providers):

Honest correction (important): "car-dog cooperation" is not a shelf product with its own page at DeepRobotics or GuoZi — it's an integrator/project-level solution combining "heavy AMR + quadruped." What you can cite officially is each maker's quadruped product page + inspection case¹³¹⁵; don't treat it as a single SKU. For Otsuka, this is value-add we integrate, not a machine off the shelf. (Note: the once-cited Rocos×Spot setup is dead — Rocos was acquired by DroneDeploy in 2021 and its site is down, so it's omitted here. ✅ Also: DeepRobotics' site deeprobotics.cn, whose cert expired 2026-06-25, is now reachable again.)

Three deployment pain points (probe each when speccing with vendors):

1. Dog-carrier docking accuracy: how does the dog board/dismount the carrier safely and precisely? Mainstream methods are AprilTag-guided self-climbing / mechanical-rail slide-in / a micro lift-ramp on the AMR — needs ≥99.9% docking success under real plant lighting and floors.
2. Multi-robot dispatch (RCS) integration: the plant scheduler usually only manages AGVs — who commands the dog now? The master must support VDA5050 or open middleware to choreograph "AMR drive → arrive → wake dog → dog climbs → dog returns → AMR departs" into one automated workflow — exactly the entry point for this project's software differentiation.
3. Cross-floor offline autonomy: if Wi-Fi/5G is poor upstairs/in a mezzanine, the dog needs strong offline obstacle avoidance + autonomous return — even after losing the link it must retrace to the staircase and re-establish connection with the carrier.

Aside · the broader "shuttle-carrier" commodity (NOT quadruped, but same "carrier-ferries-child" architecture — don't conflate):

• Dense-storage four-way shuttle-carrier (Shuttle-Carrier): a mother RGV runs the main rail to feed the child into the aisle; the child reaches deep into the racks to store/retrieve pallets — a core ASRS device, very mature and readily buyable: Dematic Multishuttle, Inform/Yinfei, DAMBACH shuttle-lift system, iAmech. ⚠️ These are wheeled shuttles that don't climb real stairs — not the same as a "quadruped child," only sharing the "mother-child" label.
• Heavy-road piggyback trucks: SANY/Shaanxi/FAW "full-out, empty-back" heavy trucks where the smaller child rides on the parent's frame home to halve tolls — SANY[source]. Unrelated to intralogistics; only explains the origin of the term "子母车 (mother-child vehicle)" in Chinese industry.
• Academic background: reviews of autonomous mobile-robot localization and AGV smart-warehousing — ScienceDirect and ResearchGate[source].

Form 3: Heavy hydraulic quadrupeds (pure-leg, but built for heavy load) — pushing payload past 100 kg 🔬 research/military

Motor-driven pure-leg quadrupeds (Unitree B2, Spot) cap out at 20–80 kg. To break "low payload" at the root, researchers brought traditional heavy-duty hydraulics to quadrupeds:

• SJTU "Baby Elephant": motor-throttled hydraulic drive + serial-parallel hybrid legs; ~130 kg self-weight, 50–100 kg payload, built for heavy hauling on complex terrain (see the CJME heavy-legged-robot review¹⁰. ⚠️ Note: the often-miscited arXiv 2202.08620 is Kirin and was withdrawn by its authors — it is not Baby Elephant; see Marsupial research review.
• Heavy quadrupeds at NUDT and other institutes: aimed at military logistics / mountain transport, with powerful hydraulic joints that shoulder 100+ kg up 30°+ slopes and rubble (cited from the user's doc; primary source to verify).
• Trade-off: hydraulics lift payload, at the cost of heavy weight, high noise/power, complex maintenance, high price — and they're almost all research/military prototypes, no intralogistics product. Not a near-term procurable option for Otsuka, but it answers "can a quadruped carry heavy?" — yes, just not on the shelf.

📊 Three-tier maturity table (buy-it-now / R&D prototype / concept-patent)

⚖️ Fusion products vs traditional wheeled AGV/AMR: differentiation & limits

Differentiation (where fusion wins):

• The only thing that climbs real stairs/steep ramps/steps — the absolute blind spot of wheeled AGV/AMR, and the entire reason fusion machines exist.
• One machine, many jobs: inspection + haulage + obstacle-crossing on the same unit; cheaper than buying a fleet of single-purpose machines when routes vary.
• Unstructured floors: gravel, gratings, ramps, thresholds — passable where a wheeled AGV is stuck.

Limits (state these plainly to the boss/client):

• Payload still an order of magnitude lower: fusion "carry-up-stairs" is still ~40 kg class; traditional AGV/AMR reach hundreds of kg to tonnes on flat ground. Heavy cross-floor loads still need the carrier route or split trips.
• High cost: industrial wheeled-legged machines cost far more than a flat-ground AMR of the same payload; heavy hydraulic quadrupeds more still.
• Low maturity: haulage use is much newer than inspection; production cases and reliability data are limited — a PoC is needed.
• Software stack not aligned with AGV standards: traditional AGV/AMR fleets interop with the master/MES via VDA5050 and similar; fusion machines mostly ship only SDK/ROS2 and rarely claim VDA5050 compliance — folding into Otsuka's existing AGV dispatch system needs integration work (each vendor's latest support to verify). This is exactly the entry point for this project's "software × hardware" differentiation: buy the hardware off the shelf, let in-house software supply the dispatch/coordination/VDA5050 bridge.

🇯🇵 Japan market / Otsuka deployment

• Procurement channel (the most concrete lead): Unitree B2-W has an official Japan distributor, TechShare (techshare.co.jp)¹¹— quote via a proper channel, with local support and 技適/certification easier to negotiate. Better suited to an Otsuka-facing commercial proposal than direct overseas import.
• Deployment cases: as of 2026-06, no publicly searchable production case of "wheeled-legged quadruped doing multi-floor intralogistics" in Japan was found (to verify); most Japanese cases are still flat AMR + elevator integration, or quadrupeds for inspection. That's an opportunity — an Otsuka PoC could be one of Japan's earliest wheeled-legged haulage deployments.
• Recommended path: first run a PoC with B2-W (via TechShare) for "cross-floor haulage + stair-climbing + in-house dispatch software"; for true heavy cross-floor loads, use the "heavy AMR + B2-W carrier relay" narrative — don't promise a single unit hauling tonnes up stairs.

🧹 Errata: two corrections to the user's research doc

The user's Google-AI-mode doc's framework (three fusion forms) is excellent and organizes this article. After cross-checking primary sources, two claims are corrected to avoid misleading the client:

1. "Unitree W1" → should be the B2-W. Unitree's industrial/logistics wheeled-legged model is the B2-W (B2 plus foot hub-motors); the official site lists no "W1" industrial wheeled-legged model — the doc's "Unitree W1" looks like a model-name mix-up, corrected here. (Unitree also has the H1/G1 humanoids and pure-leg B2/Go2 — none positioned for wheeled-legged haulage.)
2. "Direct-Drive Minimotors Tita" → wrong maker attribution. Tita is a wheeled-legged robot by Direct Drive Technology (本末科技, Shenzhen); Minimotors is a different company (the Korean e-scooter maker behind Dualtron), unrelated to Direct Drive — the doc conflated them; split here. Tita's exact walking payload (kg) still needs the official datasheet.

Also: the doc files Pudu D5 under "wheeled-legged mainstream cases." Note that Pudu officially leads the D5 as a 'delivery + inspection' quadruped, with logistics one modular app, not a dedicated warehouse haulage AMR³; Pudu's actual intralogistics flagship is the wheeled semi-humanoid D7 (not a quadruped). Don't call D5 a "warehouse AMR" when citing it.

⚠️ To verify / to add

1. Real cargo weight + route (most critical): per-item weight, cross-floor or not, step count/slope, takt — with these, the three tiers above narrow to 1–2 candidates.
2. B2-W quote / 技適 / lead time via TechShare: want me to draft a quote-request checklist for TechShare?
3. VDA5050 / dispatch integration: each fusion machine's (B2-W, Pudu, RIVR) actual VDA5050 / mainstream-AGV-master support needs written confirmation from each vendor — it decides whether they can join Otsuka's existing fleet.
4. Public commercial case for the carrier / car-dog route: confirmed DeepRobotics (Jueying X30/X20), GuoZi Robot, Unitree B2, and Boston Dynamics Spot all have a quadruped + integration path, but none has a standalone "car-dog cooperation" whole-machine SKU on its site (it's integrator/project-level). To cite a "buyable car-dog case" to the client, I should dig for a primary deployment (plant name + video + payload data).
5. Primary sources for Chinese makers: Tencent Max/Ollie, Direct-Drive Tita, NUDT heavy quadrupeds are largely second-hand citations; I can add official/paper primaries and verify specs on request.
6. Want a filterable "form × maturity" comparison table (xlsx)? Say the word and I'll generate it.