Today Gravis Robotics, the ETH Zurich spinout building autonomy for earthmoving equipment, announced a $200M investment from Softbank. This Series A marks the largest Series A in construction robotics history.
We met CEO and Co-Founder Ryan Johns in the summer of 2024 and backed the company at seed in early 2025. What we believed then has held up under contact with real job sites.
The short version: the hardest unsolved problem in construction isn’t driving a machine around a site. It’s what happens at the tip of the bucket. Gravis solved that part first, then built a product strategy patient enough for the industry that has to adopt it.
Excavation is a dexterity problem wearing a vehicle costume
Autonomous haul trucks have been running in mines for a decade, so it’s reasonable to look at an excavator and assume it’s the same problem with a smaller engine.
It isn’t. A haul truck navigates around the world along a pre-planned route. An excavator reshapes it. Every cycle, a steel bucket pushes into material it cannot see or predict, frost, then soft loam, then compacted fill, then bedrock, and the machine has to decide by feel how to keep filling without stalling the engine or pitching the chassis. That’s contact-rich manipulation, closer to a robot hand learning to grasp than a car learning to merge. And manipulation is exactly the frontier that stayed hard while perception and navigation got commoditized.
The analogy is only half right, though, because earthmoving inherits the hard parts of the self-driving stack too: localizing on a site with no lanes and no map, maintaining a 3D model of terrain you are changing several times a second, tracking trucks and people, reasoning about traversability so you don’t drive off a bench. Gravis has to be good at both. Most teams are good at one.
Classical control breaks here, and every major OEM has learned this the expensive way. They have all tried autonomous digging, and they all hit the same wall: you cannot plan a trajectory through soil whose behavior changes halfway through the cut.
Gravis learns the interaction instead. The controller trains in massive simulation across enormous variations in soil, geometry, and machine dynamics, and it uses sensing that gives it a superhuman feel. In a demo Ryan showed us early on, the same controller takes a deep clean scoop in soft soil and, seconds later, hits a hidden boulder, scrapes along the stone without rocking the machine, and still comes out with a full bucket. It never saw the rock.
That’s the part that surprises people: this isn’t imitating an expert operator, it’s beating one. A human feels interaction force secondhand, once the cab starts rocking. Full buckets and consistent cycle times are the entire productivity equation in earthmoving, and in a head-to-head trench test against an experienced operator, Gravis was ~30% faster.
Data efficiency is the whole game
If you come from autonomous vehicles, your instinct is to instrument a huge fleet, drive tens of millions of miles, and mine the tail. That playbook isn’t available here. Every hour of data costs a $250K machine, a site, an operator, fuel, and a contractor’s patience, and it isn’t portable: soil in a UK quarry doesn’t behave like soil on a South-American pipeline right-of-way.
So the approach has to be data-efficient. Gravis attacks that three ways:
Simulation as the primary teacher. I’ve spent a good part of my life figuring out how to simulate real-world processes, so I’m a believer. Gravis comes out of the lab that pioneered reinforcement learning for contact, work that became the foundation for how much of the field trains legged robots in sim, and that propagated into NVIDIA’s Isaac tooling. Making sim-to-real transfer work for interaction is a far harder discipline than making it work for navigation, and they inherited a decade of knowing how.
The right sensor for the right signal. The right sensors give the model a dense, physically meaningful signal about what’s happening at the bucket, instead of inferring it from vision. Cheaper sensing, richer supervision, faster learning.
A product that annotates its own data. Because Gravis ships assisted and augmented modes, operators intervene constantly, and the system already knows what task was underway when they did. Every intervention is a labeled example of what the machine should have done. Most of the industry’s data problem is annotation; Gravis produces perfectly annotated interventions as a byproduct of being useful. Units in the field don’t just generate revenue, they generate curriculum.
The market is bigger than the sticker price suggests
Earthmoving equipment is a $110B annual market, about a million units a year. But the number that matters is 10.4 billion machine-operating hours every year, and Gravis doesn’t need new machines to reach them. There are ~5M active earthmoving machines globally, and Gravis is compatible with a large enough share of them.
The demand pull isn’t speculative. Construction labor productivity has been negative since 2015 while manufacturing gained 18%. The US was short 350K construction workers in 2024, with 41% of the workforce reaching retirement age by 2031. Every EPC we spoke to said a version of the same thing: “ability to resource” is now a scoring question on bids. Operator scarcity is an existential constraint, not a cost line.
Why “productivity first” is the right path to full autonomy
This is where most people get the sequencing wrong, and where our diligence moved us the most.
The tempting thesis is: remove the operator, capture the salary. Companies have died on it. If you take the human out and the machine gets slower, nobody buys it, the operator’s wage is a small fraction of the cost of the hour. The machine rental, the fuel, the trucks waiting to be loaded, the grade checker, the surveyor, the foreman, the ~20% of project cost lost to rework: that’s where the money is. A 30% faster cycle pulls all of it forward.
So Gravis inverted it, and leads with productivity today. That shape wins for three reasons. It monetizes from day one instead of waiting for full autonomy to be perfect. It fails functionally, hit a boulder halfway into a trench and the system flags it, a human takes over in under a minute, and the site keeps moving. And every one of those handovers is training data.
Why we’re excited
Ryan and co-founder Dominic Jud have built the right team with the right approach, with deep roots from the ETH Zurich in the lab of Marco Hutter: a lab that has done more than any other to make learned control work on real machines, paired with people who know this industry cold. That combination is rare: roboticists who have run machines through the night on real sites, sitting next to people who know how a dealer price list works.
Construction is 12% of a $1.6 trillion pool of heavy work, it gets harder to staff every year, and the physics of the problem have finally met the maturity of the tools. The company that solves the bucket first, and earns its way to the site, wins.
We’re thrilled to have partnered with the Gravis team at seed and look forward to what’s ahead.