Monoid Labs is an AI research lab. We build small models that see, understand and act among people: on robots, on phones, and out on the street.
§ 01Why
Intelligence doesn’t happen in a datacenter. It happens at crossings , on platforms , in night markets and in the rain , twenty-four frames a second, among people who never stop moving. We build small models that keep up, and fit on a robot, a phone or a street corner.
§ 02How a model sees
A camera sees light. A model has to decide, many times a second, which of it matters. Most of our work lives in that gap: getting from millions of pixels to the few things worth acting on, with a model small enough to run where the camera is.
Layer 01 / Light
§ 03Contact sheet
Monoid 400▸ 1A
Monoid 400▸ 2A
Monoid 400▸ 3A
Monoid 400▸ 4A
Monoid 400▸ 5A
Monoid 400▸ 6A
Monoid 400▸ 7A
Monoid 400▸ 8A
Monoid 400▸ 9A
Monoid 400▸ 10A
Monoid 400▸ 11A
Monoid 400▸ 12A
Monoid 400▸ 13A
Monoid 400▸ 14A
§ 04Research
01RoboticsVision-language-action models that take in camera frames and an instruction and put out motor commands. We study how to pre-train them on web and robot data, and how to adapt them to a new body or task from a few demonstrations.
02EfficiencyModels that run on a robot's onboard computer, a laptop or a phone, and that groups without a frontier budget can study. Distillation, data selection, and architectures that keep capability as parameters get scarce.
03FoundationsOptimisation, reinforcement learning and search methods that make training cheaper, steadier and easier to reason about. We prefer methods whose assumptions can be checked, and we report the runs that fail.
04AgentsAgents that work over long horizons in real software, trained with reinforcement learning on real sessions and runnable environments. This is the home of OpenRL, our open corpus of consented agent trajectories.
§ 05Projects
OpenRL is our open, consented corpus of real coding-agent sessions and runnable environments. Agents get better by learning from real work; OpenRL lets universities, sovereign AI labs and open model builders learn from the same kind of work as the companies that run them.
Contributors use the Monoid app for macOS, Windows and Linux. It finds agent sessions on your machine, redacts them locally, and asks you to approve each one.
Shadows on the stripes · halftone
Every dot is a value
§ 06Principles
We look for the smallest model that does the job, then try to make it smaller. Small models fit on robots and phones, and more people can afford to study them.
Language is not the end of the task. We evaluate models on outcomes: code that runs, grasps that hold, crossings made safely.
Perception, language and control with clean interfaces between them, so an improvement in one part carries to the rest. That is the monoid in the name.
Every result ships with the evaluation that produced it, the baselines it beats, and the runs that failed.
Real places are crowded, badly lit and always changing. We test for that, not only for the benchmark.
Data that comes from people is collected with their consent and can be withdrawn. OpenRL was built that way from the start.
§ 07Join
If you work on robot learning, small models or RL and want to do it in the open, write to us with something you have built.
hello@monoidlabs.aiCo-design evaluations with us, use OpenRL data in your research, or tell us which environments would make it more useful.
Apply for OpenRL dataWe are seeking research grants, compute and institutional partners for open data, evaluations and releases.
partners@monoidlabs.ai
