Anatomy of a Quadruped Robot

A 2026 quadruped robot is the most engineering-mature class of legged platform you can buy. Humanoids are still arguing about hand designs and torso layouts; quadrupeds have settled. Twelve actuators, three per leg, body-mounted depth camera, top-deck for sensors or an arm, hot-swappable battery pack, Wi-Fi back to a base station. The interesting differences are no longer in the leg topology — they are in the IP rating, the payload, the spares pipeline, and the price tag.

The reference set in this article is the eight platforms a Czech buyer can actually obtain in 2026: Unitree B2, Unitree B2-W, Unitree Go2, Unitree Go2-W, DeepRobotics X30, DeepRobotics Lite3, Anybotics ANYmal D, and Boston Dynamics Spot. They span roughly $2,800 to $200,000 fully configured. That is a 70× spread for what looks like the same animal in the marketing photos. The differences matter a lot.

Leg topology — why everyone settled on 3 DoF per leg

Quadrupeds converged on the same leg design about six years ago and nobody has good reason to deviate. Each leg has three actuated joints: a hip abduction/adduction (sideways), a hip flexion/extension (forward/back), and a knee. That gives 12 actuated joints total — twelve, no more, no less, across every single platform listed above.

What does change is the foot. Three families:

• Point feet with a compliant rubber pad. The default. Cheap, replaceable, no actuation, fine on indoor floors, gravel, packed dirt, stairs.
• Wheel-feet (hybrid). Powered wheels attached to the end of each leg. Walk on rough terrain, roll on flat. Unitree Go2-W and B2-W, DeepRobotics has variants, Swiss-Mile turns ANYmal into this. Adds 4 more actuated DoF (one per wheel), so a B2-W is effectively a 16-DoF platform.
• Articulated foot. Rare on commercial units; you see it in research papers. Adds an ankle DoF for better ground contact. Not worth the wiring on a production unit unless the task is extreme terrain.

Wheel-foot is not a gimmick. On flat or paved environments it triples effective range and halves energy-per-meter. The B2-W is rated 50 km on one charge with 40 kg, where the walking B2 is rated about 20 km unloaded. For inspection routes with long corridors or a paved approach, the hybrid is the right answer.

Leg / locomotion comparison:

Two things worth noting. First, the Chinese platforms are dramatically faster than the Western industrial units. B2 hits 6 m/s, Spot tops at 1.6 m/s, ANYmal D at 1.2 m/s — not because Boston Dynamics and Anybotics cannot build a fast robot, but because their oil-and-gas / inspection customers do not want a 6 m/s machine running around live equipment. Top speed is a research metric, not an industrial one. Second, wheel-foot is a 16-DoF platform — four real motors that need power, control, feedback, and are the first thing to fail on a wheel-foot unit.

Actuators — QDD across the board

Quasi-direct-drive (QDD) is universal at this scale. Outer-rotor BLDC motor, low-ratio reducer (5:1 to 10:1), backdrivable, impact-tolerant. The humanoid article (R01) goes deeper. For quadrupeds the case is cleaner still: legs need to be light, dynamic, and impact-survivable; QDD delivers all three.

Per-joint peak torque ranges from 30 N·m on the Go2 hip to 360 N·m on the B2 knee, correlating almost exactly with platform weight. Spot and ANYmal D sit in the middle — sized for industrial payload, not record speed. Hydraulic actuators are gone from this class entirely; Boston Dynamics' Spot is electric and always has been.

On-board compute

Same two-computer split as humanoids: a real-time motor controller closing the joint loop at 500 Hz to 1 kHz, plus an application processor for perception, SLAM, and any local ML. The motor controller is rarely advertised; the application processor is.

A pattern worth naming: the Western industrial units (Spot, ANYmal) ship with a strong x86 core for autonomy software and treat GPU compute as a payload, not a chassis component. The Chinese research-grade units (Go2 EDU, B2) ship with a Jetson directly on the spine. Both approaches work; the first is friendlier to enterprise IT, the second to roboticists running PyTorch.

An on-board Jetson Orin NX (100 TOPS) handles 360° LiDAR SLAM at 10 Hz, YOLOv8-s on the head camera at 30 FPS, a 7B INT4 LLM at 15–25 tok/s, and a small VLM (Qwen2.5-VL 3B) at usable rates. It does not host a 70B LLM, a serious 72B VLM, or any training workload. That gap is what I01 is about — the quadruped does local autonomy on-board, the GPU server does the heavy thinking over Wi-Fi 6E.

Sensor suite

Every modern quadruped ships with the same shopping list, with platform-specific upgrades. The base set:

The optional extras that distinguish tiers:

• 3D LiDAR (top-mounted). Unitree 4D LiDAR L1 on Go2 EDU and B2 by default; Hesai XT16 on Go2-W Ultimate (120 m range, ±1 cm); Velodyne / Robosense / Ouster on ANYmal D and Spot as payload. The single biggest spec to read.
• Rear-facing RGB-D camera. Standard on Spot and ANYmal D, optional on B2 and X30. Matters when the robot walks backward or in a corridor.
• Thermal camera. Optional payload on X30, ANYmal D, Spot. Mandatory for electrical / mechanical inspection.
• Pan-tilt-zoom optical. Spot Cam 2 (Jan 2026) is the reference: 4K, 25× zoom, integrated radiometric thermal, 360° spherical. ANYmal D ships an equivalent.
• Foot pressure sensors. Spot and ANYmal D have explicit load cells; Unitree and DeepRobotics derive contact from IMU and joint torque. Load cells are more robust on slippery surfaces.

A Go2 EDU + L1 LiDAR is enough for indoor mapping and a research lab. For outdoor industrial inspection with 360° depth, thermal, and optical zoom you are looking at Spot or ANYmal D, and the cost reflects it.

Battery, runtime, hot-swap

Quadrupeds win here over humanoids: hot-swap battery is the norm, not the exception, and runtimes are longer because the robot is not fighting gravity to stay upright.

True hot-swap — keeping the robot running while you change the pack — is a Spot signature feature and is available on ANYmal D. Unitree and DeepRobotics call it "quick-swap" because the unit has to be in a safe pose (sitting on its hocks). For continuous-patrol deployments where you cannot afford a thirty-second downtime, only the Western units qualify. Plan for two batteries per robot at minimum, three for long shifts.

IP rating — the real outdoor-industrial divider

This is the spec that separates "research dog" from "industrial dog" more cleanly than anything else.

Two things to know. Spot's IP rating is lower than the Unitree B2's. This surprises people. Boston Dynamics built the autonomy and the spares pipeline; the chassis is rated for typical industrial weather, not monsoons. ANYmal D goes further. ANYmal X (Ex-proof) is in a class of its own for refinery work. And IP67 does not mean "swim with it" — it means dust-tight and 30-minute immersion at 1 m. Continuous rain on a moving robot with flexing seals will eventually find a way in.

Payload classes

Payload is where the price tier shows itself. The consumer / research units carry sensors. The industrial units carry sensors plus a serious arm or a custom inspection rig.

Note the sustained vs peak split. The consumer-tier numbers are roughly honest; the industrial-tier "static" numbers (B2's 120 kg standing) are real but irrelevant for any motion task — the platform cannot walk under that mass. Read the sustained walking-load number, not the headline.

Terrain handling

Marketing claims here run loose. The honest, observed capability:

What no quadruped does well: deep snow, water deeper than the foot, wet metal in winter, or terrain with overhangs the planner cannot see. Narrow spiral stairs where the tread is shorter than the foot pad eat all of them.

Communication

Wi-Fi 6E is the practical floor for off-board-compute use cases. Spot, ANYmal D, and X30 will all happily run a private LTE/5G modem instead, which is what most multi-site industrial deployments do — the robot joins a private cellular network at the customer site and inspection data syncs to a central server. Tethered Ethernet development is most common on Unitree EDU units during SDK work; industrial units assume development in simulation and wireless deployment.

Manipulation — when the dog has an arm

This is the feature that pushes a quadruped from "sensor platform" to "do something." Three platforms ship a serious arm option:

• Boston Dynamics Spot Arm. 7-DoF, 11 kg payload, ~$65k separately. Opens doors, turns valves, picks up small objects. Mature autonomy, working SDK examples.
• Unitree B2 with Z1 arm (or third-party). 6-DoF, 5 kg payload, bolts to the top deck. Cheaper but less integrated; you write the manipulation stack.
• Anybotics ANYmal D with manipulation payload. Per-deployment integration with several arm vendors for inspection tasks (buttons, valves).

DeepRobotics has shown X30 with arm payloads in research contexts but it is not a productized SKU. Unitree Go2 with arm is a research curiosity — the platform is too small to carry a useful arm. If your task involves "hand on something" interaction, you are looking at Spot. Sensor-carrying only, the cheaper platforms work fine.

Pricing tiers, Q1 2026

Approximate list / street prices in USD, ex VAT, ex shipping. Where vendors do not publish, distributor list is used. Treat as ballpark.

The honest read. Under $10k buys you a developer toy (Go2 base, Lite3 Basic). $10–30k is the research-grade sweet spot (Go2 EDU+, Go2-W, Lite3 LiDAR) — Jetson on board, decent LiDAR, real SDK. $40–90k is industrial-grade where Unitree B2 and DeepRobotics X30 live; same physical capability as the Western units at half the price, but a thinner spares and support pipeline outside Asia. Above $100k is the Spot / ANYmal tier, where you are paying for autonomy software, manipulation, and ten years of integration know-how. The price gap is real, the capability gap is real, and the support-network gap is the biggest gap of all.

What differs across platforms

A one-line map:

• Unitree Go2 — dev toy, then research platform at the EDU+ tier. The right learning robot.
• Unitree Go2-W — hybrid mobility for indoor / paved environments; cheap way to get LiDAR + wheel-foot in one package.
• Unitree B2 — industrial-grade Chinese unit, IP67, 60 kg, fastest in this list, lowest cost per kg of payload.
• Unitree B2-W — same chassis on wheel-feet, the long-range patrol option.
• DeepRobotics Lite3 — the cheapest LiDAR-capable research dog with a credible SDK and ROS 2 story.
• DeepRobotics X30 — full industrial unit, IP67, cold-weather rated, half the price of Western equivalents.
• Anybotics ANYmal D — Swiss industrial unit, deep autonomy, IP67, multi-year reference deployments in oil and gas and energy.
• Boston Dynamics Spot — the autonomy benchmark, the manipulation benchmark, the most mature SDK, and the most expensive when fully loaded.

If you want one sentence per platform: Go2 = price, B2 = speed-per-dollar, X30 = industrial-per-dollar, ANYmal = ruggedness, Spot = ecosystem.

Maintenance reality

Quadrupeds wear in a predictable order: foot pads first (cheap, quarterly), then knee QDD modules (2,000–4,000-hour service interval on dynamic use), then hip cables (vibration fatigue, one failure per quarter on heavy units), then wheel hub motors on wheel-foot variants, then battery packs (300–500 cycles before capacity loss, so 6–12 months on a hot-swap fleet). LiDAR connectors fatigue; thermal/optical payloads drift and want annual recalibration.

Vendor parts availability is the differentiator at this tier. Unitree has the most mature pipeline. DeepRobotics is improving fast in Europe. Anybotics (Zurich-based) and Boston Dynamics both have strong European support. Prague/Brno lead times: 1–2 weeks on Unitree and DeepRobotics, 2–4 weeks on Anybotics, similar on Spot.

Where Kentino fits

The same way as in the humanoid article: Kentino is a Czech channel partner / co-sell on the compute side. We do not sell quadrupeds. We build the K-AI on-prem inference servers (4× or 8× RTX 5090 or RTX Pro 6000 Blackwell, EPYC or Xeon host) that a quadruped fleet talks to over Wi-Fi 6E or private cellular. The relationship is symmetric with humanoids: the robot does its safety and local autonomy on-board, the heavy thinking lands on a GPU server, the inference server is the same regardless of whether the client is a Go2 or a Spot.

The pairing math — which inference server class fits which robot fleet — is in I01 and will be deepened in I05 (the reference build with parts list and benchmarks).

---

This is part of the Kentino Wiki, a reference series on AI compute, robotics, and the systems that connect them. Comments and corrections welcome at info@kentino.com.