One intelligence. Many bodies.
We are building a foundation model for robots: a single learned intelligence meant to control humanoids, legged machines and arms, instead of one policy per body.
Today, robot intelligence is trapped inside the body it was trained on. Change the legs, the sensors or the reach, and most of what the robot knew has to be learned again.
We think that is a data problem, not a law of nature. A model that has seen many bodies should learn what they share: contact, balance, reach, cause and effect.
Different bodies.
One model underneath.
-
Quadruped
Four legs and a trunk. A test of gait, terrain and recovery: the same physics a humanoid uses, with a different body around it.
- Actuators
- 12
- Mass
- 15.2 kg
-
Arm
Seven joints and a gripper, bolted to the floor. A test of reach, grasp and precision, where the body never has to balance at all.
- Actuators
- 8
- Mass
- 17.5 kg
Push it yourself.
This is the physics engine we train in, running in your browser. The controllers here are simple and hand-written, not our model. They show the physics a general model has to master.
Downloads about 6 MB. Runs best on a laptop or desktop.
Pretrain once.
Adapt to each body.
-
1
Describe the body
Joints, limits, sensors and mass go in as input. The model knows what it is driving before it moves.
-
2
Learn from many machines
Pretrain on simulated experience across many bodies, tasks, contacts and failures, so physical knowledge is shared instead of siloed.
-
3
Adapt, then ground it
Fine-tune for a new body with little data, then close the gap between simulation and hardware with measured real-world runs.
Proposed milestones
From a multi-body dataset to a model that transfers to a body it has never seen.
Multi-body dataset
Simulated experience across simplified bodies, tasks, contacts and failures.
Shared pretraining
One model family trained across several bodies, replacing one policy per body.
Transfer benchmark
Adaptation speed and task success measured on a held-out body.
Hardware validation
A narrow capability moved to a physical rig, with the reality gap measured.
Research stage, before a first model
We are not showing you a trained model yet.
The first proof is narrow on purpose: one shared policy that completes a measurable task across several simulated bodies. Every robot on this page is a simulation render of an open model, not hardware we have built.
The body is only the beginning.
We are looking for industrial partners with hard, multi-body problems and engineers who want to test this idea with us.
hello@superhumanoids.ai