Forward deployed engineering for robotics.
1618 Robotics is a forward deployed engineering firm based in Geneva.
We work with robotics companies and frontier labs that need embodied intelligence running outside the lab. In practice that means we collect and label the training data, source and integrate the hardware, evaluate and fine-tune the models, and place engineers on site until the system holds up without us.
Getting a robot to perform once, in a controlled setting, is no longer the hard part. Keeping it performing every day in a real facility still is, and that second problem is the one we staff for.
1.618 is the golden ratio: perfection, approached one imperfect step at a time.
Robotics today is assembled from parts that were never meant to meet: models from Californian labs, hardware from Shenzhen, data gathered wherever people happen to be.
Our work is making it converge.
Imitation learning is limited by the manipulation data you can actually get. We run collection across the three modalities that matter and handle the labeling behind them.
Egocentric video
−
Contributors wear head-mounted cameras and record everyday manipulation in their own homes, kitchens, workshops, and warehouses. The result is first-person footage of real tasks in the cluttered environments robots will eventually face, at a cost per hour that on-robot collection cannot approach. We handle contributor recruitment, task design, device logistics, and payment, and every recording is reviewed before it enters a dataset.
Teleoperation
+
When a model needs data from its own embodiment, we run teleoperated collection on the target platform. Leader-follower rigs and VR setups produce demonstrations with exact state and action pairs, recorded on the same hardware the policy will control. We build and maintain the rigs, train the operators, and manage throughput so collection keeps pace with training.
Universal gripper
+
A handheld gripper with a wrist-mounted camera lets an operator record manipulation trajectories anywhere, with no robot present. It is the fastest way to gather transferable manipulation data at scale, and it works in environments where bringing a robot would be impractical. We supply the devices, run the operator network, and deliver data in the format your training stack expects.
Labeling and QA
+
Raw episodes are not training data. Our annotation pipelines cover task segmentation, object and grasp labeling, and success tagging, with quality criteria agreed before collection starts. Episodes that fail are documented with the reason, so acceptance rates are an input you can plan around rather than a surprise at delivery.
A robotics program is also a supply chain problem. We take it from a bill of materials to a machine running on your floor, and we keep it running.
We buy directly from manufacturers in Shenzhen and across Asia, with working relationships that go back years. Arms, grippers, sensors, compute, and teleoperation rigs are specified against your requirements and budget, and we manage lead times, import, and the inevitable revisions between what the datasheet says and what arrives in the crate.
Components become a system through mechanical fit, wiring, drivers, calibration, and safety interlocks, and through the middleware nobody advertises but every deployment depends on. We do this work ourselves rather than coordinating subcontractors, which is why our timelines hold.
We install and commission on site, train the people who will operate the system, and stay through the first weeks of real use. When we leave, your team has documentation it can maintain and a system it understands.
For programs that want us to stay, we run the fleet past commissioning: monitoring, incident response, spares, and repairs, with what the field teaches feeding back into the next revision.
A policy that scores well on a benchmark can still fail on your task. We evaluate models against the work they will actually do.
We test candidate models on your tasks, your hardware, and your environment, not only on public suites. You get task-level success rates, failure mode analysis, and a clear picture of where a model is usable today and where it is not.
We fine-tune on data collected for your embodiment and your tasks, then re-evaluate against the same baseline so the improvement is measured rather than claimed. Data collection and fine-tuning run as one loop, which shortens the path from a gap in performance to the data that closes it.
Robotics is a physical business, and geography is part of the service.
GenevaHeadquarters and engineering. Central European time, an hour from most of our clients' offices, and a neutral base for programs that span suppliers and labs on three continents.
ShenzhenSourcing and hardware. Being close to the factories means specification questions are settled in days rather than shipping cycles, and quality problems are caught before a crate leaves the port.
Collection footprintContributor and operator networks across Europe, Asia, and the Americas. Data can be collected in the languages, homes, and workplaces of the markets a robot is being built for, not only in the country where the lab happens to sit.
The data and the deployments only hold up if the practices behind them do.
Every contributor is paid, every recording is made with documented consent, and licensing terms are settled before collection begins. You receive data your legal team can approve, with the paper trail to show for it.
Collection and processing follow GDPR where it applies and local data regulation where it does not. Faces, screens, and documents captured incidentally are handled according to rules agreed with you in advance.
Deployments carry risk assessments, interlocks, and operator training as standard, because a robot that hurts someone is a program that ends. Our engineers hold the same safety bar on your floor as on ours.
Client programs are segregated. Your task definitions, your data, and your results are yours, and nothing collected for one client trains anything for another.
Four kinds of teams tend to find us.
Teams whose models are data-hungry and whose robots are expensive. We widen the data funnel with egocentric and gripper collection, and keep on-robot time for the demonstrations that genuinely need it.
Labs that need embodiment-specific evaluation and fine-tuning without building a field operation of their own. We become the arms and legs of a research agenda.
Companies buying robots rather than building them, who need an engineering partner on their side of the table for selection, integration, and the first year of operation.
Makers of arms, grippers, and sensors who need their products integrated, benchmarked, and represented well in real deployments.
We work forward deployed. Our engineers embed with your team, on your site, inside your stack, and they carry the problem until it is solved rather than handing back a report.
Two to three weeks. We define the task, the acceptance criteria, and the budget with you, and we put all three in writing. If we think the problem is not ready for us, we say so here.
A bounded first deployment or collection run against the criteria from the scope. Fixed cost, fixed timeline, and a clear result: the numbers either clear the bar or they do not.
Scale is a decision you make after the pilot proves out, not a commitment you make on day one. Throughput, fleet size, and team grow against demand, with the same reporting you saw in the pilot.
We work toward leaving. Documentation, training, and tooling transfer to your team on a schedule set at the start of the scale phase, and we stay reachable after we are gone.
Geneva is home base. The work happens wherever the machines are.
Scoping starts with a note · hello@1618robotics.com1618 Robotics · Geneva, Switzerland
