Call to arms | Telekin

Weeding crops in California from your hometown in Honduras

Miguel weeds lettuce in California's Salinas Valley. He lives in a caravan shared with three other men, forty minutes from the field. He sees his kids on a phone screen twice a week when the signal holds. Most of what he earns, he spends on rent and the bare necessities; the rest goes home without him. He has done this for nine years.

Nothing about the work requires any of this. The lettuce doesn't care that Miguel left his family to tend it. The only reason a person has to uproot their life and cross a border is that, until now, the hands doing the job had to be attached to the brain of the person doing it.

That's the assumption we're breaking. Picture the same job (same crop, same field, same skill) done through a humanoid robotic avatar on the same Salinas dirt by a teleoperator 4,000 kilometres away. They walk their kids to school, put on a headset, work a shift and log off in time to eat dinner at their own table. The field gets weeded. Nobody has to live in the caravan.

We won't pretend this brings Miguel home: the workforce standing in those fields today is caught in a system we're trying to make obsolete and we'll be straight about who pockets what further down. But the choice between feeding your kids and living with them is an artefact of geography, and geography is obviated by teleoperation.

We are Telekin, two friends building teleoperated humanoid robots for agriculture (and eventually, space). This post is about why we think teleoperation, not autonomy, is the correct way to build useful robots this decade, what our prototype has taught us thus far and what we're doing next.

Autonomy is a data problem and nobody has enough

Billions of dollars are being incinerated by some very intelligent people waiting for physical AGI to wake up. The autonomous humanoid gold rush is real, however most of the frontrunners will fail, not because the hardware is hard (it certainly is) but because of a far simpler problem: their AI brains need decades' worth of multimodal, embodied training data, yet none of them are collecting it at a meaningful rate. Benchmark GP, Eric Vishria, has benchmarked the industry as having one millionth of the data it needs. We think that's generous. A human child absorbs 80,000 hours of multisensory experience before being let loose on the world and starts the clock with millions of years of evolutionary pretraining baked in.

Tesla crossed a comparable gap for full self driving by putting a fleet on the road that people paid for while it quietly collected data. Autonomous driving, whilst undeniably impressive, is a highly constrained problem: a planar space, well-defined rules, painted lines and standardised signage. A humanoid in a home or factory has orders of magnitude more degrees of freedom and faces vastly more varied challenges yet there is no equivalent data collection fleet in the wild today.

The industry's current answer is paying gig workers to film themselves repeatedly opening microwaves and folding towels. The companies hoovering up this footage are burning cash prodigiously with not a cent of revenue in sight. Data collected as a by-product of real, paid, quality-controlled work is categorically better and arrives with revenue attached rather than a bill.

Teleoperation is how you build that fleet now without waiting for AGI to show up. A human mind is a bona fide general intelligence, already trained for the physical world. We are not short of humans on this planet; our bodies are simply unoptimally distributed. Minds, it turns out, can now be redistributed at the speed of light. Every hour a teleoperator works is a chargeable hour of revenue and an hour of embodied training data. The robot pays for its own education. Better still: a company that owns the whole vertical, from robot design through to field operations, is collecting data perfectly kinematically matched to its own humanoids. No retargeting losses, no simulation gap. Every paid hour widens the lead. The moat digs itself.

What a one-armed prototype taught us

We are early. Earlier than most companies would be comfortable admitting. We have a scratch-built, one-armed semi-humanoid proof of concept called Horatio. He lives in a nest of wires, his off-the-shelf joints have emitted so much magic smoke that anyone would think we were hotboxing the office, yet he's taught us far more than a prettier robot ever could. We've used him to test our core assumptions, remotely pack a lunchbox, grade the odd tuber, raise a li'l bit of capital and not much else. That's rather the point of this post - we're recruiting!

For our lunchbox packing test we gave Horatio three omniwheels and the run of our kitchen. He gracefully picked up a satsuma from the fruit bowl and retrieved a drink from the fridge. Then a cupboard door exposed a critical difference between our compliant meatbag arms and his stiff aluminium manipulator: it...