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

⚠️ This is a draft to verify. Each "quote" is pulled word-for-word from a vendor's official page (quoting is for verification, not original authorship), and the small superscript at the end of each sentence is clickable and jumps to that sentence's source page. But price, Japan procurement channels, and certification (Giteki / Radio Law) are unverified; a few figures are not given on the official page (e.g. DEEP X30's page lists no payload), each flagged "not on official page." The source for Unitree B2/Go2 has been switched from the shop.unitree.com storefront marketing pages to the www.unitree.com spec pages (both CN and EN), and B2's IP67 and working temperature have been verified and corrected accordingly. Please verify and edit.

⚠️ Note: DEEP Robotics' official site deeprobotics.cn had its TLS certificate expire on 2026-06-25 (yesterday); the related links in this piece are temporarily unreachable and will recover once the cert is renewed — DEEP figures are treated as "to verify" meanwhile.

🌐 What this piece answers

The boss's real question isn't "what capabilities do robot dogs have," but "given these capabilities, what's worth building." So this piece doesn't stop at a capability list — it places robot dogs on a set of unified 11 capability axes to paint a strengths-and-weaknesses profile, then translates that profile into a task judgment.

The 11 capability axes (shared across all three pieces in this group, for easy side-by-side comparison): ① mobility ② payload ③ manipulation dexterity ④ perception ⑤ autonomy (SLAM · task autonomy) ⑥ human-robot interaction ⑦ runtime ⑧ secondary-development openness ⑨ reliability maturity ⑩ cost ⑪ certification (outdoor · explosion-proof etc.).

Every judgment is backed by a verbatim quote from the official page, with a source superscript at the end of the sentence. For robot-dog vendor comparison / price / selection see the sibling Robot-dog products; for other form factors (humanoid / wheeled / arm / dexterous hand / wheel-leg / drone) see Other embodied-AI types & capabilities; for a quick feasibility read on specific scenarios see Candidate entry scenarios · feasibility quick-read.

Source scope (all official domains): Unitree Go2, Unitree B2, Boston Dynamics Spot and Spot Arm, DEEP Robotics X30/Lite3, ANYbotics ANYmal and ANYmal X, Ghost Robotics Vision 60.

📊 11 capability axes: robot-dog strengths-and-weaknesses overview

This is a one-page profile compressed from the evidence below. Ratings are the typical level of robot dogs as a class (not a single model); the superscript on each evidence sentence opens for verification.

At a glance: robot dogs are nearly all green on ①②④⑤⑧⑩⑪ (mobility · payload · perception · autonomy · dev · cost · certification), red only on ③ manipulation dexterity and ⑥ human-robot interaction — and those two are precisely the humanoid's home turf. The evidence unfolds axis by axis below.

🐾 Capability deep-dive (capability × representative product × official quote × source)

1. Mobility / terrain (walk, stairs, slopes, speed)

This is the robot dog's strongest, most mature capability.

• Unitree B2 — among industrial quadrupeds it leads on speed and obstacle-crossing: the official page says "Fastest running industrial-grade quadruped robot at 6m/s"; obstacle-crossing "Superior obstacle-crossing ability, easy to cross the mess of wood piles, 40cm high platforms".²
• DEEP Robotics X30 — stair-climbing: the official page says it can "flexibly climb up and down stairs at a 45° angle" and "stably climb industrial stairs" (spec: speed ≥4 m/s, slope ≤45°, obstacle ≥20cm).⁶
• Boston Dynamics Spot — official spec: Max Speed 1.6 m/s, Max Slope ±30°, Max Step Height 300mm, and "self-righting if it falls".⁴
• ANYbotics ANYmal — complex terrain: the official page says "Safe on open grated stairs", "Works on wet and bumpy terrain", walking speed "0.75 m/s".⁸
• Ghost Robotics Vision 60 — stresses legs over wheels/tracks: "Legs outperform wheeled and tracked robotics", with "Robust recovery and inverted operation in case of slips or falls".¹⁰

2. Autonomous navigation / perception (SLAM, obstacle avoidance, GPS-denied)

• Ghost Robotics Vision 60 — GPS-denied autonomy is the standout: "fully autonomous operation in both GPS-enabled and GPS-denied environments", paired with "real-time 3D lidar-based SLAM mapping".¹⁰
• Boston Dynamics Spot — inspection-grade autonomy: "autonomously charging, dynamically replanning around new obstacles, and self-righting if it falls", with "360° perception and athletic intelligence".⁴
• ANYbotics ANYmal — built for human-made multi-floor plants: "AI-based mobility for navigating complex, multi-floor plants made for humans", standard 360° LiDAR + thermal + ultrasonic mic.⁸
• DEEP Robotics X30 — extreme-environment autonomy: "integrated perception, it can autonomously navigate and operate in extreme environments", and can work in "darkness, strong light, flickering, and even without any light source".⁶
• Unitree Go2 — even the entry model carries LiDAR: "Standard Ultra-wide 4D LiDAR upgrades Recognition system by 200%", min detection distance "as low as 0.05m".¹

3. Payload / mounting (as a mobile platform carrying loads)

Key point: payload is the "carry sensors" class, not the "exert large force to work" class — expanded under "ceiling & limits."

4. Environmental tolerance (dust/water / temperature / explosion-proof)

• ANYbotics ANYmal X — the benchmark in this category: explosion-proof certification "ATEX & IECEx certified up to Zone 1 IIB", "Certified for up to Zone 1 where inflammable gases are likely to occur", plus "IP67: Water and dust ingress protection". (The page notes "2026 specifications coming soon" — the full spec table isn't yet public.)⁹
• DEEP Robotics X30 — "IP67 protection", operating temperature "Operate from -20°C to 55°C", "Unstoppable operation in Extreme weather".⁶
• Ghost Robotics Vision 60 — "IP67 rating, for operation in the most extreme weather", temp range "-40° to 55°C".¹⁰
• ANYbotics ANYmal — "dust and waterproof (IP67 rated)".⁸
• Boston Dynamics Spot — IP54, -20°C to 55°C (from the official spec sheet / support article, not the landing-page marketing copy).¹¹
• Unitree B2 — ✅ the official spec pages (both CN and EN) confirm "IP67" and working temperature "-20℃ to 55℃" (note: the shop.unitree.com storefront marketing page indeed lists no IP, but the www.unitree.com spec page does — verified and corrected accordingly).³

5. Runtime (active time per charge)

6. Secondary development / SDK openness (this is exactly the key to "platformization")

• DEEP Robotics Lite3 — the most direct openness statement: "Provide models, motion control SDK and APIs, as well as perception development interfaces with sample code."⁷
• Ghost Robotics Vision 60 — rich interfaces: "C/C++, ROS, ROS2, MAVLink Compatible, Zeno, ATAK, JSON Mission", onboard compute "NVIDIA® Xavier 32GB RAM".¹⁰
• Boston Dynamics Spot — "robust API, and comprehensive documentation", plus "Available software API and hardware interface control document".⁴
• ANYbotics ANYmal — onboard "2× Intel i7 Core", built-in "WIFI & 4G/LTE".⁸
• Unitree Go2 — X/EDU models support "Secondary development" and "Graphical programme".¹

7. Manipulation (an arm — the quadruped's weak spot, requires an add-on)

• Boston Dynamics Spot Arm — the strongest in this category, yet the official wording is still "constrained": "Grasp, lift, carry, place, and drag a variety of items with the arm's 6-degrees of freedom and gripper."; "Semi-autonomously turn valves, flip levers, open doors, and manipulate other objects with constrained movement."; "Lift up to 11kg and drag up to 25kg with the arm."⁵
• Ghost Robotics Vision 60 — "Manipulator Arm" is listed only as an optional payload module (the capability exists; no detailed spec on the main page).¹⁰
• Unitree / DEEP / ANYbotics — the quadruped pages above all carry no arm / manipulation claim (they are perception + mobility platforms; manipulation requires buying an arm separately).

🟢 What's possible now (backed by official quotes)

• Robust legged locomotion across human-made and rugged terrain: 45° stairs (X30), 300mm steps (Spot), 40cm platforms (B2), wet/grated stairs (ANYmal), self-righting (Spot, Vision 60).
• Autonomous inspection & navigation, incl. GPS-denied: ANYmal "navigating complex, multi-floor plants made for humans"; Vision 60 "fully autonomous… GPS-denied" + "3D lidar-based SLAM"; Spot "dynamically replanning around new obstacles".
• Sensor-carrying mobile platform: meaningful payload (B2 >40kg, Spot 14kg, ANYmal +10kg, Vision 60 10kg) + integrated LiDAR/thermal/acoustic sensing.
• Harsh / hazardous-environment operation: X30, ANYmal, Vision 60 generally IP67; Vision 60 -40 to 55°C; ANYmal X uniquely explosion-proof "ATEX & IECEx… Zone 1 IIB".
• Constrained single-arm manipulation: Spot Arm can semi-autonomously "turn valves, flip levers, open doors", "Lift up to 11kg".

🟡 What could grow with secondary development

• Open SDK/ROS stacks → custom autonomy and reinforcement learning: Lite3 "SDK and APIs… with sample code", Vision 60 "C/C++, ROS, ROS2", Spot "robust API", Go2 EDU "Secondary development". This is the main growth axis: custom perception, mission logic, learned gaits.
• Modular mounting bays → new missions: ANYmal "Payload expandable" (can mount gas sensors etc.), Vision 60 swappable modules (incl. "Manipulator Arm"), Spot's T-slot/DB25. Stack domain sensors (gas/radiation/RF) without swapping hardware.
• Onboard compute already runs heavier models: Vision 60 (NVIDIA Xavier 32GB), ANYmal (2× Intel i7) — headroom to run custom vision/LLM-assisted autonomy on-device.
• Manipulation is the frontier add-on: arms already exist (Spot, Vision 60), extensible with better grippers/learned policies — but still "constrained movement" today.

🔴 Capability ceiling & limits (vs humanoids)

The evidence is mainly the absence of a corresponding official claim + the "constrained" wording where manipulation is claimed:

1. Dexterous/general manipulation is weak or absent. Most quadruped pages (Go2/B2/X30/Lite3/ANYmal) carry no manipulation claim at all — they are perception + mobility platforms. Even the strongest, Spot Arm, is officially limited to "constrained movement", a single 6-DoF arm with a single gripper; not one vendor claims two-handed dexterous assembly or tool use.
2. Not built for human workstations / ergonomics (low interaction approachability). Quadrupeds sit low (Spot step 300mm; ~0.4–0.5m standing). The value prop is "navigating complex, multi-floor plants made for humans" — traversing human spaces to "inspect," not standing at a bench to "work" at human waist/eye height or to "greet and interact" with people. A single ~1m-reach arm (Spot) only partly closes this; humanoids are designed around the human workspace and tool layout (see Other embodied-AI types & capabilities).
3. Limited single-charge endurance. Inspection-grade ~90 min (Spot, ANYmal) to ~3 h (X30, Vision 60) of continuous walking — short for continuous shifts, mitigated only by autonomous docking.
4. Payload is platform-class, not work-force class. B2 (>40kg carry) and Spot (14kg) are "carrying sensors," not exerting large manipulation forces; Spot Arm tops out at "Lift up to 11kg".
5. Marketing-vs-spec gaps (flagged honestly): DEEP X30 official page lists no payload; ANYmal X "2026 specifications coming soon"; Spot's IP54/runtime/temperature come from the spec sheet, not the landing page. (Note: Unitree B2's IP was previously mis-flagged as "not on the official page" — its shop.unitree.com storefront marketing page indeed has none, but the www.unitree.com spec page lists IP67; verified and corrected.)

🧭 Task matching: what these capabilities suit / don't suit

This is the landing point of the piece — translating the 11-axis profile above into the judgment "should you use a robot dog," instead of stopping at a capability table.

✅ The robot dog's home turf (tasks its strengths map to directly)

❌ Tasks the robot dog shouldn't be forced into (red-zone weak spots)

💡 One-line insight (form factor matches task)

A robot dog = an excellent autonomous mobile sensing platform. For any task that is "move + sense + outdoor/hazardous + run a fixed route," its 11-axis profile is nearly all green, its cost-performance and maturity are both solid, and it can be secondarily developed for differentiation — inspection / transport / data collection are its home turf. For any task that is "two-handed operation of human tools" or "human-facing interaction," it falls into its ③⑥ red zones and should be handed to a humanoid or a wheeled robot with a screen. The boss's intuition — "robot dogs are cheap, start by using them as a platform" — is right for inspection/transport/data-collection scenarios; whether to do it, and which one, see Candidate entry scenarios · feasibility quick-read.

⚠️ For you to verify / to add

1. Price/procurement/certification: no quotes above; Japan channels, warranty, and Giteki are in the sibling Robot-dog products and still to verify.
2. Not-on-official-page items: DEEP X30 payload — defer to the formal datasheet/quote. (Unitree B2's IP67 / working temperature have been verified on the www.unitree.com spec page and are no longer listed as missing.)
3. Whether to turn the 11 axes into a scorable comparison table (capability × product, cell-by-cell scoring), or add more models like LimX, new Unitree, etc. — just tell me how to organize it.