Cambrian

We’re a small group of engineers and roboticists exploring ideas that are newly within reach. Most of us are Stanford researchers who also work in labs like ILIAD, REAL, and Hazy Research.

We’re starting now because the economics and the possibilities of robotics research are shifting.

1. Running an experiment used to take weeks: integrating components from several vendors, calibrating poses, and fabricating a test rig, to name a few. Frontier AI has collapsed the cost of software and is beginning to do the same for parts of hardware.¹ Most of the setup is no longer the bottleneck.
2. Frontier LLMs are also demonstrating striking spatial reasoning capabilities, strong enough to generate realistic 3D environments and even actuate robots directly.² We think much of the field still underestimates how much value these models can unlock when used creatively.

Together, these mean that testing a strange idea now costs a fraction of what it did even a few months ago. Directions that once needed a well-funded lab and years of infrastructure are now within reach of a small group with a good idea and the will to pursue it.

The name comes from the Cambrian explosion, when a convergence of environmental factors let life go from simple organisms to most of the animal forms alive today over a short stretch of geological time. We think robotics is nearing a similar moment. The right factors are lining up for an explosion in both deployment and new form factors, built to serve all sorts of functions. We’re eager to help bring it about, with the resources to take the work seriously and the freedom that comes from working outside commercial pressure.

Early directions

Right now, we’re most excited about three directions. The first is unexplored end effectors and dexterous tool use. The field has mostly converged on parallel-jaw grippers and dexterous hands,³ and we think the space of possible bodies, and how one collects data for them, is much larger than what’s been tried.

The second is automating the novelty pump. More data doesn’t automatically make a robot policy better; it only improves from data that shows it something new, like a situation where it fails or a task it hasn’t tried. Today, collecting that data is manual work: people watch the robot, spot where it struggles, and set up new demonstrations, and that effort doesn’t compound as the robot is deployed more. We think frontier models can take over parts of this loop, from spotting failures to generating new scenarios. The goal is a true data flywheel, where every hour of deployment makes the policy better, which is harder than it sounds.

The third is what we call embodied alignment: making sure robots act in ways that follow our intent, including the parts we don’t think to specify.⁴

Reach out

If you have a bet you want to test and Cambrian sounds appealing to you, we’d like to hear from you, whether or not you’re at Stanford or in a lab. Labs and companies interested in sponsoring research or learning more about what we’re up to are welcome too.

Email us at mmusic@stanford.edu.