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

⚠️ This is a draft to verify. Each "quote" is pulled verbatim from a vendor's official page, and the footnote at the end of each sentence is clickable and jumps to that sentence's source page. All Tesla official/IR domains return HTTP 403 (blocked) to the fetcher, so Tesla Optimus and its dexterous hand are flagged UNVERIFIED / to verify — no official quote obtained. The direct robot-dog-vs-humanoid comparison has no official source whatsoever and is flagged "analysis (to verify)."

🌐 What this piece answers

Robot dogs (quadrupeds) are just one form factor of embodied AI (Physical AI). To decide "what's worth building," the other form factors must be put on the same yardstick: which forms can do what robot dogs can't? On which exact axes does the boss's concern — "the humanoid's advantage in intelligence and form-factor fit" — actually land?

This piece uses the exact same 11 capability axes as the Robot-dog capability map: ① mobility ② payload ③ manipulation dexterity ④ perception ⑤ autonomy ⑥ human-robot interaction ⑦ runtime ⑧ secondary-dev openness ⑨ reliability/maturity ⑩ cost ⑪ certification. First a six-form horizontal matrix, then each form factor backed by official quotes + footnotes, finally a head-on answer to the boss's core question landing on task matching. For concrete scenarios see Candidate entry scenarios · feasibility quick-scan.

📊 11 capability axes: six-form horizontal portrait matrix

⚠️ This is an analysis (to-verify) relative portrait: the glyph indicates that form's relative strength on that axis (🟢strong / 🟡medium / 🔴weak / — not applicable), not official cell-by-cell data. Line-by-line evidence is in each form's section below (footnotes jump to official pages). On the cost axis 🟢=relatively cheap, 🔴=expensive.

How to read this table: each form's "green zone" is its home turf. Quadrupeds are green on mobility · perception · autonomy · cost · certification; the humanoid's only exclusive greens are ③manipulation dexterity + ⑥human-robot interaction (plus generalist AI on top); wheeled is green on interaction · runtime · maturity · cost but all-red on ③manipulation; the robot arm is green on manipulation · payload but all-red on ①mobility; the drone is green on aerial mobility · autonomy but all-red on manipulation. There is no all-around form factor — selection is just "which green zone does the task fall into."

🤖 Category 1: Humanoid

Humanoids are the form factor that most thoroughly matches the human world, and the most aggressive on the AI-learning path — they exclusively own the two green zones ③manipulation dexterity + ⑥human-robot interaction.

• Figure 03 — AI as the prerequisite for scale: officially "There's no path to scaling humanoid robots without AI.", pursuing "large-scale, end-to-end pixels-to-action learning" (Helix model); dexterous manipulation "fine-grained, dexterous control over fragile, irregular, or moving objects"; fingertip touch "can detect forces as small as three grams of pressure".¹
• Boston Dynamics Atlas — fleet-level skill transfer + same-workstation work: "when one Atlas learns a new skill, that task can easily be deployed across your entire Atlas fleet."; "Atlas is made to operate within the same workstations using the same equipment your staff does"; self battery-swap "autonomously navigates to a charging station and swaps out its own battery". Official spec: 56 DoF, 50kg peak/30kg sustained payload, reach 2.3m, 4h runtime, IP67.⁴
• 1X NEO — home + generalist model + remote fallback: "NEO arrives with basic autonomy for early owners and grows in capability overtime"; "NEO uses Redwood AI—1X's Generalist AI model—for learning and repeating tasks"; long-tail tasks "an Expert from 1X can remotely supervise its actions".⁵
• Agility Digit — already commercial + no site retrofit needed: "A fully autonomous tool with proven commercial deployments"; "human-centric form factor means that manufacturing floors and warehouse spaces don't have to be redesigned"; "Designed to excel in spaces where people already work".⁶
• Unitree G1 — force-control dexterous hand + learning-driven: "Force control dexterous hand, manipulation of all things"; "Imitation & reinforcement learning driven" (UnifoLM model). Spec: 23 DoF (EDU 23–43), ~35kg, single-arm payload ~2–3kg, price US $13.5K.²
• Unitree H1 — speed record: "Moving speed of 3.3m/s(world record), Potential mobility > 5m/s".³
• Tesla Optimus — ⚠️ UNVERIFIED: tesla.com and IR domains all return 403; no official quote obtained. The widely circulated "general-purpose humanoid for factories/homes" phrasing is a search-engine paraphrase of the official page, not a reliable verbatim quote → to verify.

Today: material handling is in commercial deployment (Digit official blog "Digit Moves Over 100,000 Totes in Commercial Deployment"); autonomous + automatic battery-swap (Atlas); end-to-end AI learning as the main line (Figure pixels-to-action, 1X Redwood, Unitree UnifoLM); payload from light (G1 ~2kg) to industrial-grade (Atlas 50kg peak, NEO official "Lift 154 lbs").

Secondary-dev growth: official SDK/API; interchangeable end effectors (Digit "interchangeable end effectors"); expandable DoF (G1 23→43); model-driven iteration + remote-expert fallback for the long tail (NEO).

Ceiling & limits: runtime is generally ~2–4h, continuous work depends on battery-swap/fast-charge; light platforms carry little; home units are currently "basic autonomy + remote human fallback" (NEO says so officially); most official pages don't publish full hardware specs; whole-unit cost is far above quadrupeds.

🛒 Category 2: Wheeled Service

• Pudu BellaBot — indoor autonomous delivery: navigation "Industry-exclusive Dual SLAM Solution Covering All Scenarios"; avoidance "3D Omnidirectional Obstacle Avoidance", response "as short as 0.5 seconds"; ops "Quick battery swap makes it easy for BellaBot to serve 24/7". Spec: max payload 40kg (10kg/tray), speed 0.5–1.2 m/s, slope ≤5°.⁷
• Keenon — partly verified: the homepage claims "World's First Self-Developed VLA Model for the Service Industry" (referring to the KOM 2.0 model); but the T10/W3/T6 detail pages failed to fetch, hard specs to verify.

Today: flat-floor indoor point-to-point autonomous delivery; tens-of-kg multi-tray payload (BellaBot 40kg); sub-second 3D avoidance; multimodal voice/touch interaction; hot-swap battery 24/7.

Secondary-dev growth: elevator/access/POS/scheduling API integration; advertising & VLA task models (Keenon KOM 2.0); custom tray accessories; multi-robot dispatch.

Ceiling & limits: no arm — cannot grasp/manipulate, goods must be loaded/unloaded by hand; flat floors only (slope ≤5–7°, no stairs); no outdoor/explosion-proof certification; "intelligence" is navigation + scripted interaction, not general manipulation. But ⑥human-robot interaction + ⑦runtime + ⑨maturity + ⑩cost are all green — the most cost-effective form for pure-interaction/delivery scenarios.

🔧 Category 3: Robot Arm

• Universal Robots UR20 — heavy-payload collaborative arm: "longest reach of our heavy payload cobots"; "UR20 is our fastest cobot"; quick to start "the first simple task is typically less than an hour". Spec: payload 25kg, reach 1750mm, weight 64kg.⁸
• Franka Research 3 — force-sensitive dexterous research arm: "The 7 degrees of freedom Arm of Franka Research 3 offers human-like dexterity"; joint torque "The torque sensors integrated at each joint finely estimate external contact forces"; real-time control "FCI provides low-level, real-time control of the robot". Spec: 7 DoF, payload 3kg, reach 855mm, repeatability <±0.1mm, control 1kHz.⁹

Today: precise repeatable manipulation within a fixed workspace (pick/place, machine tending, palletizing, welding, assembly). Heavy cobot (UR20 25kg/1750mm); high-DoF force-sensitive arm (FR3 7DoF/1kHz) for contact tasks and robot-learning research.

Secondary-dev growth: UR extends via URCaps/grippers/vision; Franka via FCI real-time control + ROS for imitation/RL and force-control research; both extend through end effectors + sensing + policy learning.

Ceiling & limits: fixed base — no mobility (①all-red); workspace bounded by reach; payload and dexterity trade off (UR20 25kg but industrial 6-DoF; FR3 dexterous/force-sensitive but only 3kg/short arm). It manipulates within reach, it does not move through the environment.

🤲 Category 4: Dexterous Hand

• Shadow Dexterous Hand — anthropomorphic five fingers: "The most advanced 5-fingered robotic hand in the world"; "in-hand manipulation", "Tendon Driven"; serves as "testing hardware for AI and Machine Learning". Spec: "20 actuated DOF and a further 4 under-actuated movements for a total of 24 joints", "over 100 sensors running at up to 1KHz". Its RL variant DEX-EE is officially "Designed and built for the world's leading AI research team" (Google DeepMind).¹⁰
• Tesla Optimus hand — ⚠️ UNVERIFIED / to verify: tesla.com/AI returns 403. The widely circulated "22 DoF/hand, tendon-driven, forearm actuators" figures are Musk/team X statements relayed by third parties, not an official fetchable page.

Today: anthropomorphic five-finger grasp/pinch/in-hand re-orientation; fingertip force/position control ~1kHz; dense touch; RL/AI manipulation research hardware (Shadow's 20–24 joints approach the human hand's richness).

Secondary-dev growth: ROS integration → stack learning policies (RL/imitation), tactile-driven manipulation, teleop data collection to train foundation models; Tesla's tendon-driven route points to manufacturable, low-cost dexterity.

Ceiling & limits: research hands (Shadow) are expensive, maintenance-heavy, low-payload, tethered to a controller — not field-rugged; humanoid hands trade DoF for manufacturability, with touch/force/durability still below the human hand; truly reliable general manipulation is still an open frontier. It is a component of the "manipulation dimension," not a standalone platform — most other axes are not applicable.

🛞 Category 5: Wheeled-Legged

• Unitree B2-W — wheel-leg hybrid. ✅ Unitree's official CN/EN spec pages now publish the full numeric spec (correcting the earlier "no spec on the official page" call): continuous walking load >40kg, standing peak load 120kg, wheel speed 15km/h (wheel 50rad/s), step 20–25cm, climb up/down 40cm, slope >45°, range 25km at a 40kg load, total weight ≈85kg (incl. ~12kg battery), IP67, working temperature −20℃ to 55℃. Note: the widely cited "120kg" is the standing peak — the continuous walking load is the same >40kg as the legged B2; the wheeled-legged gain over the B2 is flat-ground speed and range, not heavier carry-up-stairs payload.¹¹
• ANYbotics / Swiss-Mile — ⚠️ UNVERIFIED: no citable official spec fetched; wheel-speed/runtime/efficiency figures are all third-party reports, to verify.

Today: hybrid mobility — fast/efficient wheeling on flat ground, lock wheels to cross steps/obstacles; tens-of-kg payload; inspection/security/logistics; some can stand and use front legs as an arm.

Secondary-dev growth: mount arm/LiDAR/sensors; autonomous inspection routes; last-mile delivery; fleet ops.

Ceiling & limits: wheels need fairly traversable terrain, extreme rubble/soft ground still favors pure legs; payload/battery vs agility trade off; high cost (the CN official site routes to a sales inquiry, the overseas store lists ~$100k); the wheeled-legged "carry-up-stairs payload" is no larger than pure legs (continuous walking >40kg) — don't mistake the 120kg standing peak for hauling capacity.

🚁 Category 6: Drone

• DJI Matrice 400 — long-endurance heavy-payload inspection: avoidance "integrated rotating LiDAR and mmWave radar for power-line-level obstacle sensing"; "a payload capacity of up to 6 kg"; "an impressive 59-minute flight time"; smart flight "Real-Time Terrain Follow, Cruise Mode, Smart Track, and POI". Spec: transmission 40km (FCC), max level speed 25 m/s.¹³
• Skydio X10 — onboard-AI full autonomy: "Backed by an onboard NVIDIA Jetson Orin GPU, the X10 harnesses unrivaled computing power to make the right decisions in real-time"; "Six custom-designed navigation lenses provide 360-degree visibility"; tracking "Skydio Shadow enables seamless tracking of people and vehicles"; night flight "operate 24/7 in NightSense mode for zero-light navigation". Spec: 40 min flight, max level 20 m/s, IP55.¹⁴

Today: long endurance (DJI ~59min, Skydio 40min); modular heavy payload (DJI 6kg multi-gimbal); onboard-AI full-autonomy avoidance + target tracking; long-range transmission; night/all-weather inspection.

Secondary-dev growth: SDK + modular mounts → custom sensors, automated inspection missions, dock/fleet autonomy, AI data analytics, BVLOS; Skydio edge-autonomy supports GPS-denied/indoor.

Ceiling & limits: payload-vs-endurance trade-off, lifting work is capped; battery limits mission time; regulation (BVLOS/airspace) and weather limit deployment; DJI faces geopolitical/procurement limits in some markets; autonomy in cluttered/adversarial environments is still imperfect.

🎯 Core conclusion: limits of robot dogs vs humanoid AI's advantage in "intelligence × form-factor fit" + task → form

Official quotes supporting the humanoid advantage (verbatim, each with a footnote)

• Same workstation, same equipment (no machine-specific retrofit) — Atlas: "Atlas is made to operate within the same workstations using the same equipment your staff does".⁴
• Human-centric form → no site redesign needed — Digit: "human-centric form factor means that manufacturing floors and warehouse spaces don't have to be redesigned".⁶
• Designed for spaces where people already work — Digit: "Designed to excel in spaces where people already work".⁶
• Generalist AI is the prerequisite for scale — Figure: "There's no path to scaling humanoid robots without AI.".¹
• Generalist model drives learning — 1X: "NEO uses Redwood AI—1X's Generalist AI model—for learning and repeating tasks".⁵
• One learns, whole fleet deploys (humanoid fleet intelligence) — Atlas: "when one Atlas learns a new skill, that task can easily be deployed across your entire Atlas fleet.".⁴
• Anthropomorphic dexterous hand (a manipulation dimension the dog form lacks) — Shadow: "The most advanced 5-fingered robotic hand in the world".¹⁰

Form-factor comparison points (analysis, to verify — not vendor comparison quotes)

⚠️ No official page directly writes a comparison sentence like "robot dogs are worse than humanoids." Below are inferences based on the verified quotes above, flagged as analysis (to verify).

• Robot dogs are strong on mobility, weak on manipulation (③⑥red): quadruped official pages lead with mobility/payload/runtime; their value is a "mobile platform"; to "do work" you must add an arm/dexterous hand (see Robot-dog capability map §7).
• Wheeled service robots = pure mobility + scripted interaction, structurally no manipulation: BellaBot et al. have no arm; goods must be loaded/unloaded by hand — but ⑥interaction friendliness is strong.
• The humanoid's unique combination = bipedal fit to human spaces + two-arm/anthropomorphic-hand manipulation + generalist AI learning: jointly supported by three groups of official quotes from Atlas/Digit (same workstation, no site redesign) + Shadow/Optimus hands (multi-DoF manipulation) + Figure/1X (generalist models). This is exactly the source of the "intelligence × form-factor fit" advantage: the human world is built to human scale and human-hand operation, so humanoids fit it natively; robot dogs can only traverse it to inspect, and struggle to "do the hand-work humans do" within it.

💡 Task → most cost-effective form factor (one-line decision table)

Wrap-up: there is no all-around form factor; selection = placing the task into someone's green zone. If what the boss wants to enter is inspection/hauling/data collection, the quadruped is currently the most cost-effective, most mature form, open to secondary dev for differentiation; the humanoid instead bets on "two-handed general work at human workstations" — bigger payoff but short runtime, high cost, and earlier-stage maturity. The trade-off between the two paths depends on the scenario and time window your company wants to enter — for the concrete accounting of the three candidate scenarios, see Candidate entry scenarios · feasibility quick-scan.

⚠️ Sourcing status / for you to verify

• Fully verified (official text + spec): Figure 03, Unitree G1/H1, Unitree B2-W (spec now filled in, cross-verified against Unitree's CN + EN official spec pages), Atlas, 1X NEO, Digit, Pudu BellaBot, UR20, Franka FR3, Shadow Hand/DEX-EE, DJI Matrice 400, Skydio X10.
• Partly verified: Keenon (homepage only, detail pages not fetched).
• UNVERIFIED / to verify: Tesla Optimus (whole unit), Tesla Optimus hand (all tesla.com/IR domains 403); Swiss-Mile/ANYbotics wheeled-legged (third-party reports only).
• Matrix glyphs and the robot-dog-vs-humanoid comparison sentences: all are analysis (to verify), with no official cell-by-cell source.
• If you want me to add Tesla's official quotes (playwright-rendered fetch), turn the matrix into a sortable scoring comparison table, or do form-factor selection for a specific scenario, just tell me.