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

⚠️ This is a draft under continuous verification. The specs for Unitree Go2 / B2 / B2-W have been cross-checked line-by-line against Unitree's Chinese official site (unitree.com/cn) + overseas English official site (unitree.com) — figures carrying a [source] superscript are the wording from the official spec pages (note: shop.unitree.com is a storefront page with price only and no spec table; specs always defer to www.unitree.com). The DEEP series remains marked "to be verified" because its site is temporarily unreachable. Prices are public-channel reference values, and the Japan procurement channel, certification (Giteki), and the real weight of Otsuka Shokai's goods are all unknown/to be verified. DEEP Robotics' site deeprobotics.cn had its TLS certificate expire on 2026-06-25 and is temporarily unreachable (deeprobotics.com is a parked domain, not official), so DEEP figures here carry no external link and are uniformly marked to-be-verified, to be re-linked once the site recovers.

🌐 What this piece answers (Otsuka Shokai's real need)

Otsuka Shokai currently uses AGV / AMR (automated guided vehicles / autonomous mobile robots) to haul goods inside the plant, but these wheeled vehicles can't climb stairs, can't climb ramps, can't cross thresholds — multi-floor lines with ramps/door sills are exactly their blind spot. The boss wants to use robot dogs to cover that blind spot, and once a plan is ready, pitch a partnership aimed at Otsuka Shokai's real need.

But there's a key unknown: how much each item weighs. Payload directly determines which unit to pick — and even whether "a robot dog can do it at all." So this piece doesn't rush to recommend one unit; instead it sorts robot dogs by payload tier and lays their "walking payload × stair-climbing × speed × runtime × price" into a fill-in-the-blank selection table — once the real weight arrives, look it up and slot it in, ready for whatever changes come.

For the full capability picture of robot dogs, see the Robot-Dog Capability Atlas; for a cross-vendor price review, see the Robot-Dog Product Survey; this piece is the selection view that focuses them on the single scenario of "haulage."

📊 First, get this clear: AGV/AMR vs robot dogs (payload × terrain)

This is the premise of the whole piece — the two are complements, not replacements:

In one line: Don't treat the robot dog as a "stronger AGV" and try to out-payload it — it'll lose. Where the robot dog earns its keep is the cross-floor / up-and-down-ramp / over-the-step "last leg" that AGVs can't do. Position it as a complement to AGV/AMR (flat floor to the AGV, stairs and ramps to the robot dog) and the narrative holds up.

🐾 Robot-dog haulage capability master table (by payload tier × stair-climbing × price)

"Walking payload" = the effective load it carries while walking (the figure that actually matters for haulage); "standing payload" is only the peak it can hold while standing still, which is not the same as carrying it up stairs. Figures with a superscript are official wording; the rest are to be verified.

Key reading of the table (important correction): the "continuous walking load" of both the B2 and the B2-W is officially >40kg — the frequently cited "120kg" is the standing peak (what it can hold while standing still), not what it can carry while walking, let alone carry up stairs (see the "load + stair-climbing stability" section below). The wheeled-legged B2-W's real advantage over the pure-legged B2 is faster flat-ground wheeling (15km/h) and longer range (25km at a 40kg load), not heavier carry-up-stairs payload. Pure-legged quadrupeds' walking payload generally caps at ~40kg; anything heavier needs splitting into multiple trips or an "AGV + robot-dog hybrid." Light models like Go2/Lite3 carry only a few kg, suited to small items.

🎯 Payload-tier selection (handling "unknown goods weight")

Once you have the per-item weight of Otsuka Shokai's goods, slot it straight in:

Light load ≤ 10kg (small parts, trays, documents, samples)

• First choice: Unitree Go2 EDU (~$3,790) or DEEP Lite3. Cheapest, and still climbs stairs and ramps, ideal as the first PoC unit to validate "haulage + stair-climbing + in-house dispatch software." Failure cost is extremely low.

Mid load 10–40kg (most material boxes / tote-bin class)

• First choice: Unitree B2 (continuous walking >40kg, >6m/s, step 20–25cm, up/down 40cm, slope >45°, IP67) — the best payload-and-obstacle value, the workhorse model for robot-dog haulage. If the same route is mostly flat-floor long-distance and you want more speed and efficiency: the wheeled-legged B2-W (same >40kg walking, but 15km/h wheeling, 25km range at a 40kg load).
• For more mature autonomy/certification, with ample budget: Spot (14kg) / ANYmal (+10kg) / X30 (~40kg · to verify), but at $70k–150k per unit, this leans toward "buying a finished product."

Heavy load > 40kg (pallets, full cases, heavy material)

• No single robot dog (including the wheeled-legged B2-W) can carry it up stairs — the continuous walking load of both the pure-legged B2 and the wheeled-legged B2-W is officially only >40kg; the 120kg is just a standing peak (see next section). Three ways out:
  1. Split + multiple trips: break the heavy goods into ≤40kg loads carried in several runs (sacrificing cycle time);
  2. AGV + robot-dog hybrid: keep flat-floor heavy goods on the AGV/AMR, and hand only the "cross-floor/up-and-down-ramp/over-the-step" leg to the robot dog as a relay — this also fits the "complement" narrative best;
  3. Switch to another type: a tracked electric stair-climbing cart can carry 200–420kg (but needs a human operator); a fixed route can use a goods lift / VRC — see the sibling piece Beyond the Robot Dog for Stair Haulage.

Decision tree in one line: goods ≤10kg → Go2 EDU to test the waters; 10–40kg → B2 as the workhorse (wheeled-legged B2-W for flat-floor long-distance); >40kg → split into multiple trips / "AGV runs the flat floor + robot dog takes the stairs and ramps" / switch to a tracked cart or goods lift (see the sibling piece). Get the real goods weight first, then slot it in.

🛞 Wheeled-Legged: faster and farther on flat ground, but "carry-up-stairs payload" is no bigger than pure legs

When the same route has both flat-floor long distances and occasional steps/ramps, wheeled-legged is the realistic compromise — but first dispel a common misconception: wheeled-legged does not mean it can carry more up stairs.

• Unitree B2-W: adds hub-motor wheels to the B2. The official CN/EN spec pages now publish the full numbers: continuous walking load >40kg (same as the pure-legged B2), standing peak 120kg, wheel speed 15km/h, step 20–25cm, up/down 40cm, slope >45°, 25km range at a 40kg load, total weight ≈85kg (incl. ~12kg battery), IP67⁴. Its gain over the B2 is in flat-ground speed and range, not carry-up-stairs payload.
• DEEP LYNX M20: wheel-leg hybrid, public-channel continuous payload around 15kg (to verify).
• Trade-off: on flat floor it's fast on wheels and energy-efficient; at steps/obstacles it locks the wheels and uses them as legs to cross. But on extreme gravel/soft ground/steep stairs it's still inferior to pure legs, and its full-load stair-climbing ability is on par with pure legs (look at walking load, not the standing peak), so before deployment be sure to verify against the official datasheet + an on-site test.

🔧 Load + stair-climbing: is it actually stable? How to judge it

This is what the boss cares about most, and the point most easily skewed by vendor spin: when a robot dog carries goods up stairs, will it tip over / will it be unable to carry? The answer can't rest on a single big "120kg payload" number. Below is a judgment method you can take straight to your supplier, to the spec sheet, and to on-site acceptance.

① The first key: read the "continuous walking load," not the "standing peak load"

These two numbers differ greatly in a vendor spec, and mean completely different things — haulage and stair-climbing only count the former:

Why the walking number is much smaller: while standing, the weight is supported roughly vertically by the four legs (passive load-bearing); the moment it walks, every step has a leg off the ground, the center of mass shifts, and acceleration/deceleration inertia spikes joint-torque demand, so the usable load drops sharply. Unitree officially lists these two numbers separately (B2 / B2-W: continuous walking >40kg vs standing peak ≥120kg) — that separation is itself a reminder to buyers not to order against the standing number.

② Step height / slope: always ask "was it rated under load?"

A vendor's "max step height" and "max slope" are almost always measured no-load, on dry, clean, hard ground. Under load the center of mass rises and the joint-torque margin is eaten up, so the actual passable step/slope drops. So treat the spec-sheet numbers as discounted, and ask the supplier in writing: was this step/slope measured no-load or at rated load?

③ Load it "low + centered": the center of mass decides whether it tips on a slope

Same vehicle, same slope — the higher and more off-center the load, the more easily it tips. The rule of thumb: the tolerable slope angle roughly satisfies tan(θ) ≈ (half base-width ÷ CoG height) — i.e. the wider the base and the lower the center of mass, the steeper the slope it can climb. Stacking the load high → raises the CoG → shrinks the critical slope angle → tips earlier. Using the vendor's load mounting area/interface (e.g. Spot publishes an 850×240×270mm mounting area⁶ to place the load low and centered is the cheapest stability gain there is.

④ Three engineering concepts worth understanding (layman's version)

• Support polygon: the smallest convex polygon connecting all the grounded feet's contact points. If the CoG's ground projection falls inside the polygon → it won't fall; at the edge/outside → it tips. With four feet down the polygon is large and most stable; while climbing stairs often only 2–3 feet are down, the polygon shrinks, and the stability margin drops.
• ZMP (Zero Moment Point): a special point on the ground where the horizontal tipping moment from gravity + inertia is zero. As long as the ZMP stays inside the support polygon, dynamic walking won't tip — it is the "moving version of the center of mass," the core stability criterion for legged robots.
• Why a load weakens stair-climbing: ① the CoG rises/shifts, more easily exiting the already-shrunken support polygon; ② lifting a leg onto a step is torque-heavy to begin with, and the extra load takes another slice, leaving less margin to cope with slips/missteps; ③ under high current the battery voltage sags (battery sag), further reducing peak available torque.

The above are the general engineering principles of legged-robot stability (support polygon / ZMP, introduced by Vukobratović in 1968, an industry-standard criterion); for a specific model's "how much load at how high a step without tipping," no public data can substitute for an on-site test.

⑤ Joint reading of the B2-W official numbers (the honest conclusion for the customer)

Unitree B2-W official CN/EN spec pages: continuous walking >40kg, standing peak 120kg, continuous stair-climbing step 20–25cm, single step up/down 40cm, slope >45°, 25km range at a 40kg load⁴. Reading them together:

1. The weight it can carry up stairs is the 40kg (walking) number, not 120kg. Carrying 120kg up stairs is not achievable — this must be stated to the customer up front.
2. "40kg walking," "25cm continuous stair-climbing," "40cm single step," ">45° slope" are each rated independently and can't simply be added together. There is no public data proving "simultaneously full-load 40kg + continuous 25cm stair-climbing" holds — this is exactly the point to have the supplier test on-site.
3. 25km range @40kg at least shows 40kg is a sustainable working condition rather than instantaneous; but that range is most likely measured on flat floor — climbing stairs/ramps markedly raises energy use and shortens the real range.

Cross-comparison (the B2-W is in the first tier for heavy-load stair-climbing, but inspection models are simply a different payload class):

Spot / ANYmal are positioned for inspection (light sensor loads), while the B2-W includes industrial haulage — which is exactly the basis for arguing to Otsuka Shokai that "haulage takes a B2/B2-W class, inspection takes a Spot class."

⑥ On-site test checklist (go through it item by item at acceptance / negotiation)

1. Require a "loaded stair-climbing" test: continuous up-and-down at the customer's real step height + real goods weight — don't accept a no-load or reduced-weight demo. This is the single most important item in the whole piece.
2. Order against the walking figure: set it by "continuous walking load" (B2/B2-W = >40kg); treat the 120kg standing figure only as a reference margin.
3. Discount the rated slope/step: the actual passable figure under load is usually below the spec sheet — keep margin, don't plan routes right at the limit.
4. Load it low and centered: lower the CoG, reduce off-center loading — directly improves resistance to tipping on slopes.
5. Ask in writing: were the step/slope specs rated under load or no-load? How much joint/hub torque margin is left to cope with slips and missteps?
6. Verify loaded range: ask for a range estimate that includes the stair/ramp-climbing condition, and schedule shifts and charging points accordingly.
7. Test real-floor traction: run a loaded ramp test on the customer's actual floor (which may have oil/water/metal grating), not the clean showroom floor.
8. Ask for the failure mode: the critical slope angle / step height at full load, and whether it triggers a protective stop or simply tips over when exceeded.

🧭 Advice for the boss / for pitching a partnership to Otsuka Shokai

1. Narrative positioning: cover the AGV/AMR blind spot, not replace it. Otsuka Shokai already has flat-floor AGV/AMR; the pain point is "can't climb stairs/ramps." Sell the robot dog as "relay haulage across floors · ramps · steps" — it hits the pain point squarely without fighting the existing AGVs for the same job, making it easier to pitch.
2. Ask 5 numbers first, then set the plan: ① the goods' per-item weight and dimensions ② whether it's cross-floor ③ how many steps / how steep the ramp on the route ④ cycle time (how many trips per hour) ⑤ existing AGV payload/model. Once these five numbers arrive, this table outputs the selection directly.
3. PoC path: start with Go2 EDU (light) or B2 (mid-load workhorse) to validate "haulage + stair-climbing + in-house dispatch/recognition software"; the differentiation is the in-house software (path dispatch, coordination with AGVs, goods recognition) — the hardware is bought off the shelf. This ties into the business plan's "software × hardware" combined advantage.
4. Honest risks: heavy goods (>40kg/pallets) cannot be carried up stairs by a single robot dog (including the wheeled-legged B2-W) — don't over-promise. For those, go the "AGV + robot-dog hybrid," "split into multiple trips," or switch to a tracked cart / goods lift (see Beyond the Robot Dog for Stair Haulage), and tell the customer the boundary honestly.

⚠️ For you to verify / supplement

1. Real goods weight (most critical): give me the per-item weight range of Otsuka Shokai's goods and I'll narrow the selection table down to 1–2 candidate models + a rough cost.
2. DEEP Robotics data: the official site deeprobotics.cn had its certificate expire 2026-06-25 and is temporarily unreachable — payload, stair-climbing, and price for X20/X30/Lite3/LYNX are to be re-linked and verified once the site recovers. Also: the DEEP external links in the existing Robot-Dog Capability Atlas and Robot-Dog Product Survey also point to that expired domain — tell me if you want me to change/annotate those too.
3. ✅ Standing vs walking payload (verified): Unitree's CN/EN official spec pages confirm — the continuous walking load of both the B2 and the B2-W is >40kg, and the standing peak is ≥120kg; this piece has been corrected accordingly (the earlier reading that treated 120kg as the B2-W's haulage payload was a misreading).
4. B2-W loaded testing: although the official spec is now complete, there is no public measured data for "full-load 40kg + continuous 25cm stair-climbing"; before deployment still get an official quote + an on-site loaded test (see the checklist above). DEEP LYNX's official spec is still missing, pending the site's recovery.
5. Whether to build a weight-filterable Excel/comparison table (input goods weight → auto-suggest candidate models): tell me and I can generate an xlsx selection table from the existing data.