General-purpose manipulation for embodied AI.
Building the perception and control layer that lets robots act in the physical world.
Otolit is an independent robotics research effort focused on general-purpose manipulation, the open problem at the center of embodied AI.
Locomotion is largely solved. Perception is advancing fast. But dexterous, contact-rich manipulation across diverse tasks and objects remains the bottleneck between today's robots and genuinely useful ones. That gap is where we work.
Our focus areas:
- Vision-Language-Action (VLA) models: fine-tuning and extending open foundation models for real manipulation tasks.
- Tactile sensing and control: integrating touch into the action loop, where vision alone fails for contact-rich tasks.
- Sim-to-real transfer: training in simulation and deploying on physical hardware, with a bias toward affordable, reproducible setups.
An otolith is the small mineralized particle in the inner ear of vertebrates, it tells an organism where it is in space. When you move, it shifts, and that signal becomes orientation. It is one of the simplest, oldest, most universal organs of embodied awareness in nature.
We took the name because before a machine can manipulate, navigate, or act, it has to know where it is. Embodied intelligence begins with a sense of the physical.
This organization hosts our open research and tooling:
- Fine-tuned VLA model checkpoints and training recipes
- Demonstration datasets in LeRobot format
- Simulation environments and sim-to-real pipelines
- Notes, experiments, and reproducible benchmarks
Most of our model weights and datasets live on Hugging Face.
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 build on the open tools of the community.
Otolit is based in Valencia, Spain.
We're early. If you're working on manipulation, embodied AI, or sim-to-real and want to compare notes, reach out.