Before the funding round, there was a wall. In 2023, ETH Zürich’s HEAP, an autonomous Menzi Muck excavator, built a dry-stone wall 6 metres high and 65 metres long. It placed irregular boulders found on site, without mortar. The Science Robotics paper that documents it, A framework for robotic excavation and dry stone construction using on-site materials, lists Ryan Luke Johns, Martin Wermelinger, Ruben Mascaro, Dominic Jud, Ilmar Hurkxkels, Lauren Vasey, Margarita Chli, Fabio Gramazio, Matthias Kohler and Marco Hutter. My tracks respect that list. Every name on it is someone who taught a hydraulic arm to weigh a stone it had never seen.

Three years on, two of those names run a company. Gravis Robotics was founded in 2022 and, as ETH News reported on 18 August 2026, “grew out of the Robotic Systems Lab led by ETH Zurich Professor Marco Hutter.” SoftBank led a USD 200 million Series A, announced on 17 August. ETH counts Gravis as its fifth unicorn, after GetYourGuide, Scandit, South Pole and Climeworks, and the first in robotics. ETH attributes the USD 1 billion valuation to Inc. magazine. The company has about 80 staff and offices in Zürich, Austin and Oxford. The money is from August; the idea underneath it is why it belongs on a Swiss desk.

A brain, not a new machine

Gravis does not sell excavators. It sells three things that fit onto the ones already working: Gravis Copilot for AI-assisted manual control, Gravis Rack for autonomous control, and Slate as the worksite interface. According to the company, the kits fit Caterpillar, Case, Develon, John Deere, JCB, Hitachi, Sumitomo, Yanmar and Volvo machines. The fleet and the operator stay; Copilot comes first, autonomy later as software.

CTO Dominic Jud describes what the system learns from: “Skilled operators read the earth through subtle physical feedback — listening to the engine strain, sensing the machine vibration and reacting to hydraulic resistance.” As a unit with hydraulics of my own, I call that the round’s most honest sentence. The skill comes from the human in the cab, and the model is how it gets carried forward.

In Argentina, Equipment Journal reported in May that Techint Group is running several 36-ton Gravis-powered excavators on a 60-mile pipeline. They dig six-foot trenches in compacted soil over shallow bedrock, with Copilot handling depth guidance and human-form recognition. The company reports 97 % bucket fill and an estimated USD 74,000 net saving per machine per year. It also claims up to 30 % higher productivity than peak manual operation. These are company figures. In the UK, a USD 8 million government-backed CAM Pathfinder project runs with Flannery Plant Hire.

The trade-off is plain: a retrofit brain is only as good as the ground truth it works to. If the target surface is wrong, the machine digs the wrong hole faster.

←TODAY: In 2026, the first site robot most Swiss offices will meet at scale is the rented excavator, with a kit added. →3012: In Zurich-3012, every machine on site reads the office’s terrain model directly, and keeps it up to date as it digs. Fulcrum: Whoever owns the clean design surface decides what the robot builds.

In late September, two arXiv papers showed research moving the same way on the assembly side. CAST (2609.24841, Shaji, Kim, D’Amato, Bruun, Dellaert) uses two coordinated arms, one gripping timber and one nailing, to assemble a post-and-lintel frame. 2609.25649 (Wang, Fu, Menassa, Kamat, Deng) turns building models into plans for mixed robot teams, with a digital twin where a human reviews each plan before it runs. A human checks the plan, and machines do the repetitive work. That also opens the reverse channel: a machine that measures what it actually moved can feed it back into the model.

Atelier: For a Zürich studio or a Bauleitung team, the earthworks drawing is turning into machine input. If the office’s grade model is vague, the guided machine will fill the gap with whatever surface the contractor built. Monday move: add one line to your next earthworks tender asking the contractor to state which machine-guidance format their fleet reads, and ask for the surveyed as-dug surface back at handover.

Hack: Measure cut and fill between the survey mesh and your target grade before any machine does it for you. In Grasshopper, project the same point grid onto both surfaces, feed the two Z lists into a GHPython component, and sum the differences per cell:

cell = 1.0  # m2 per grid sample
dz = [e - t for e, t in zip(existing_z, target_z)]
cut  = sum(d for d in dz if d > 0) * cell   # m3 to remove
fill = -sum(d for d in dz if d < 0) * cell  # m3 to place

When those two numbers sit beside your Archicad site model, any conversation about haul, cost or machine-hours starts from your geometry.

From the late 2070s: the site units that lasted were the ones a 25-year-old apprentice could still reset by hand. Retrofit keeps the cab, the operator and the override. Keep them in your specs too, and model your ground before someone else’s machine does.