Quadruped robots have converged on a single silhouette. Watch three promotional videos back to back and you will see the same four legs, the same confident trot across gravel, the same recovery after a shove. What the footage does not show is that two machines which look identical can be separated by an order of magnitude in what they will actually survive on a working site. 

The specification sheet does show it, if you know which lines decide the outcome. This is a guide to reading one, written from manufacturers' own published figures. 

The first question is not which brand, it is which class 

Broadly, the market splits into two families that happen to share a body plan. 

The first is the consumer, education and research platform. It is affordable, capable, quick to get moving, and it exists so that people can develop against a legged robot without a six-figure budget. The second is the industrial platform, specified for a job that runs on a schedule in weather that does not cooperate. 

Almost every disappointed deployment I have seen traces back to buying the first and expecting the second. The specifications below are where that difference actually lives.

Ingress protection tells you more than any other line

If you read only one figure, read the IP rating. It governs whether the robot can be outdoors at all, and it is the cleanest dividing line between the two families. 

The pattern in the published data is stark. DEEP Robotics rates the X30 at IP67.

Unitree rates the B2 at IP67. ANYbotics rates ANYmal at IP67. Boston Dynamics rates Spot at IP54. Unitree's Go2 does not publish an ingress rating at all, and the user documentation describes it as not waterproof and not dustproof. 

Those are not shades of the same thing. IP67 means dust tight and able to survive temporary immersion. IP54 means limited dust ingress and protection from splashing water, which is a genuine outdoor rating but not a wash-down or heavy-rain one. 

No published rating means the manufacturer is not making a claim, and you should not make one on their behalf. 

A practical test when reading any spec sheet: if the ingress rating is absent rather than low, treat the machine as an indoor and fair-weather platform regardless of what the video showed.

Operating temperature, and the phrase that should stop you

Ingress protection has a companion figure that is easy to skim past. The industrial platforms in this class publish a range of roughly minus 20 to 55 degrees Celsius. The Go2's published range is 5 to 35 degrees, qualified as good weather. 

If you are specifying a robot for a site in northern Europe, Canada or the American midwest, a 5 degree lower bound rules the machine out of most of the year before you consider anything else. Watch also for conditional wording. Boston Dynamics, for example, publishes an operating range down to minus 20 but notes that the robot should be powered on at zero degrees or above. That is a real operational constraint for a robot living in an unheated enclosure, and it is the kind of detail that only appears if you read past the headline number.

Stairs, steps and slopes are not one specification

Legged robots are bought largely for ground that wheels cannot manage, so this section deserves care. Three separate figures matter and they are often conflated: 

- Step height, the vertical obstacle the robot can clear in one movement.

- Stair capability, which is about repeated, rhythmic ascent and descent.

- Maximum slope, usually given in degrees.

Published figures illustrate the spread. The Go2 lists a step of about 15 to 16 centimetres and slopes of 30 to 40 degrees. The X30 lists a 20 centimetre step with stairs and slopes at 45 degrees. The B2 lists stairs of 20 to 25 centimetres and slopes beyond 45 degrees. 

Two cautions. First, a standard commercial stair riser in many countries exceeds 16 centimetres, so a robot rated near that figure may manage your test stairs and fail your plant stairs. Second, open grated stairways and steel walkways are a different problem from solid steps, because the robot's perception has to cope with seeing through the tread. ANYbotics calls this out explicitly for ANYmal.

Most specification sheets do not mention it at all, which is not the same as handling it.

Endurance is three numbers, not one

Runtime on a spec sheet is usually the unloaded figure, and it is the least useful version of the number. What determines whether a robot covers your round is the combination of: 

1. Runtime under load, with your actual payload fitted.

2. Range, which some manufacturers publish alongside runtime and which is    often the more honest figure. The X30 lists at least 10 kilometres, the B2    over 20 kilometres unloaded.

3. The recharge story. This is where platforms diverge most sharply. Some    list autonomous docking as standard, some as an option, and some offer only a    manual charging pile on selected variants. A robot that cannot return to a dock    by itself is not doing unattended patrol, whatever its runtime says.

Payload interacts with all three, and payload figures are themselves layered. Manufacturers often publish a continuous walking payload and a separate maximum or standing figure. The B2's published continuous walking payload of over 40 kilograms is a different class of capability from the Go2's 7 to 8 kilograms, and it is the continuous figure that governs what sensors you can actually fit and still finish the round.

Perception is where the price hides

Two robots can list similar locomotion figures and differ completely in what they perceive. Read this section for what is standard versus what is an accessory. 

At the developed end, ANYbotics publishes a payload set including 360 degree LiDAR, six depth cameras, zoom and thermal cameras, a spotlight and an ultrasonic microphone for leak detection. Boston Dynamics publishes 360 degree stereo sensing with thermal and pan-tilt-zoom available through an add-on camera payload. At the accessible end, a base platform may ship with a basic LiDAR intended for obstacle avoidance rather than mapping, with the better sensor reserved for a higher variant. 

That last distinction matters more than it sounds. A LiDAR sized for not walking into things is not necessarily a LiDAR that will build a usable map or support an autonomy stack. If your application depends on mapping, confirm which sensor is fitted to the exact variant you are quoted, not to the product family. 

Development access is frequently tiered

If you plan to integrate the robot with anything, check the software access model before you check anything else. It is commonly tiered by variant in a way that is easy to miss.

Unitree's SDK, for example, spans C++, Python and ROS 2, but access differs by model: the education variant has full access, an intermediate variant has basic access, and entry variants have none. DEEP Robotics publishes a C++ navigation SDK with ROS 2 messages. Boston Dynamics publishes a Python SDK and gRPC API.

Buying a platform for development and receiving the variant without SDK access is a costly and entirely avoidable mistake. 

What is not on the spec sheet

Three things decide industrial deployments and rarely appear in the comparison table. 

Certifications. Some manufacturers publish design references against machinery safety standards, radio compliance, and in some cases explosion-protection certification for hazardous areas. Others publish none. For a regulated site this is not a nice-to-have, and absence of published certification is a procurement problem rather than a detail.

Support and spares. A legged robot has many actuated joints. Ask what a joint module costs, how long a repair takes, and where the work is done. A cheap robot with a twelve week repair turnaround is not cheap.

Integration. Most buyers are not buying a robot, they are buying detections routed into a system they already run. Ask early how alerts and video reach your existing control room or VMS, because the answer frequently involves more engineering than the robot itself.

Testing beats specification reading

Everything above narrows the field. None of it tells you whether the machine works on your site, because published figures are laboratory and best-case numbers by nature. They are honest and useful, and they are not a test result from your yard in November.

The reliable way to close that gap is a bounded pilot: one route, run repeatedly, with the measures agreed before the robot arrives. Useful measures are coverage against the intended round, false alarm rate and what each false alarm costs in operator time, uptime including charging behaviour, and the hours actually saved.

Include an exit condition. A pilot that concludes a robot is wrong for the site has saved considerably more than one that flatters it. 

Read the specification sheet to decide what is worth testing. Then test it.

Simon Bumford is founder of Forge Robotics (https://forgerobotics.co.uk), a manufacturer-neutral UK market-entry partner for humanoid and quadruped robotics. Forge does not sell robots. It matches platforms to tasks across suppliers and designs the demonstrations and pilots that prove them, working with overseas manufacturers entering the UK market and with UK businesses evaluating adoption.

All figures quoted in this article are the manufacturers' own published specifications at the time of writing and are not Forge test results.