General-purpose manipulation for embodied AI.
We build the perception and control layer that lets robots act in the physical world, since dexterous, contact-rich manipulation is still the bottleneck between today's machines and genuinely useful ones.
Locomotion is largely solved. Manipulation is not.
Robots can walk, balance, and navigate. But picking up an unfamiliar object, handling something soft or deformable, or completing a multi-step physical task remains unreliable. That gap is where we work, on the models and the sensing that turn intention into precise physical action.
Vision-Language-Action
Fine-tuning and extending open VLA foundation models for real manipulation tasks, conditioned on natural language.
Tactile sensing
Bringing touch into the action loop, where vision alone fails. Contact, force, slip, and texture for dexterous control.
Sim-to-real
Training in simulation and deploying on physical hardware, with a bias toward affordable, reproducible setups.
Notes from the workbench.
Research notes, experiments, and findings as we build. Informal, technical, and published as we go.
Papers & technical reports.
Before a machine can act, it has to know where it is.
otolith /ˈoʊtəlɪθ/ n.
The small mineralized particle in the inner ear of vertebrates. When the body
moves, it shifts, and that signal becomes balance, orientation, a sense of
place in space.
One of the oldest and most universal organs of embodied awareness in nature.
We took the name because embodied intelligence begins with a sense of the physical.
On the shoulders of the open ecosystem.
We build on the work of the open robotics and ML community (LeRobot, Isaac Lab, ROS 2, and the open VLA model families). We publish research and reproducible results where we can, and contribute back in kind.
Our models and datasets live on Hugging Face; our code and tooling on GitHub.