News & insights

I've spent more than twenty years building robots for environments where failure can be catastrophic, from the ExoMars rover programme to surgical robotics. When I started thinking seriously about construction, I assumed it would be a slightly easier problem. In many ways, it has turned out to be harder.

Mars is unknown, but once you've landed, the terrain in front of you doesn't change that much. A construction site can evolve by the hour as materials arrive, crews move around and floors, walls and stairs gradually appear. The robot itself is part of that process too. Every time it completes an operation, it changes the environment it has to navigate next.

Then there is the physical nature of the work. Construction means lifting, pushing, drilling, fastening and manipulating heavy objects, often with considerable force and fine precision. It isn't enough for a robot to navigate successfully through a site. It has to be able to do useful work when it gets there.

That combination has shaped almost every decision we've made about the All3 Mantis.

Why we didn't build a humanoid

There's enormous excitement around humanoid robots right now, and for some environments the human form makes a lot of sense. Homes, offices and established factories have been designed for humans. Door handles, shelves, stairs and worktops are all positioned with the dimensions and capabilities of the human body in mind.

A construction site is different. It is the environment before the human environment exists. It is ‘pre-human.’ Floors may be incomplete, walls aren't there yet, and materials are often sitting in the routes you need to travel through.

That led us to a fairly basic question: if the environment hasn't been designed around the human body, why should the machine be?

I've seen this approach throughout my career. A Mars rover is shaped by its terrain, gravity, and power budget, while a surgical robot is shaped by anatomy and how a surgeon needs to work. In neither case would you start with a particular body shape and then look for jobs it might be able to perform.

Force is particularly important in construction. When you drive a structural screw into timber, every newton you put into the structure comes back into the robot. If the robot moves, the tool moves with it, and you lose precision at exactly the point where you need it.

This is one reason All3 Mantis weighs 135kg. We didn't reluctantly end up with that mass; it gives us the stability the work requires. Four legs allow us to keep that mass relatively low and distribute it effectively, while still giving the robot the mobility it needs around a changing site.

The same logic determined the rest of the machine. All3 Mantis is designed to carry 100kg anywhere on site because that is the scale of real construction components. It can pass through standard doorways, travel in construction lifts and be transported on a Europallet. We also separated the machine's height from its working reach, so All3 Mantis doesn't need to be four metres tall to work four metres above the ground. The arm goes to the work instead.

Where AI helps, and where it doesn't

In our journey to create All3 Mantis we've deliberately questioned another assumption in robotics. Right now, a lot of attention is understandably focused on what AI and learned models can make robots do. Our starting point is slightly different because we're trying to put a 135kg machine onto a real building site and have it work reliably every day.

At this scale, you have to begin with the physics.

All3 Mantis therefore has several layers of control, with a hard boundary between learning and the motors. At the lowest level, a dedicated real-time controller runs 1,000 times a second, keeping the robot stable, limiting forces and deciding what happens if anything above it stops responding.

Above that, we're pragmatic. We use AI where it gives better results and conventional control where it doesn't. Walking is an area where learning works extremely well, so we train All3 Mantis's gait in simulation using measured models of our own actuators. Fastening is different. For that task, a classical controller currently gives us the precision we need at the tool tip.

That boundary has already been tested for real. When All3 Mantis lost power at the top of a flight of stairs, it lowered itself to the ground under control, using behaviour that sits below the line, where nothing learned can reach it.

What All3 Mantis has now proved

Stairs are the most common hard obstacle on a site. All3 Mantis has autonomously climbed a 35 degree staircase, continuously up and down, carrying its arm and navigating using only its own sensing. At 135kg, and 180kg with its arm, as far as we can establish, it is the heaviest electric-drive quadruped publicly documented doing so.

That benchmark itself isn't the point. What it shows is that a robot with the mass needed for construction work can still move autonomously through a building.

More importantly, it can do the work when it arrives. On a real building component, All3 Mantis autonomously located, aligned and drove structural screws into eight holes out of eight, repeating the operation across multiple runs without an operator and maintaining precision to within 0.5mm at the tool tip.

The whole chain now runs end to end. The building model tells our Mission Control system what needs to happen, All3 Mantis travels to the location and performs the operation, and the result comes back into the system as a digital record of what has actually been built.

Fastening is the first operation, rather than the destination. We've identified 68 operations of construction work that the platform is being developed to perform.

Next comes the test that matters most: taking All3 Mantis onto a live construction site, starting in early 2027, and proving that these results hold when we don't control everything happening around them. Every decision behind All3 Mantis was made for the site. That is where it will be judged.