EIDON AI

After 2+ years, we're winding down Eidon. We started out chasing multimodal data but at the beginning of 2025 we turned toward robotics, and from there our team built a full arm-tracking and egocentric video data platform and the first Eidon hand. None of this could have been done without the people who collected, tested, argued in Discord, and backed us early. Contributors, supporters, investors, frens — thank you. It was a good run.

🌅🤖 JUST PASSED 1000 HRS OF AI ROBOTICS DATA COLLECTION (synchronized arm tracking + egocentric video). We continue to scale to 100k+hrs and include our GLOVE tech for dexterity. The full sashimi SET: glove + arm trackers + egocentric video + Sym (simulation env). 🍣 2026 is peak of the AI Robotics data collection race. DM for sample data and godspeed to 1M hrs

the trenches in AI robotics data collection in 2025-26

🎉👏 now imagine adding raw synchronised arm tracking data to this not just from factories.

IX/ If robotics is AI's endgame, a decentralized version is essential — thousands of hours of synchronized data from diverse contributors will power it. We've started with our wearable IMU trackers, capturing upper-body kinematics in real environments. There is so much more to do.

VIII/ Next: extending our data collection with Eidon Gloves and Glasses to capture fingers and dexterity data too. There is a long road of solving practical problems tho.

VII/ Eidon prototyped the full stack: Firmware for low-latency streaming, app for field capture, backend for visualization and simulation. Our long term strategy is building the whole vertical stack: data collection/post-processing, and an infra platform for custom datasets, models, simulations, and synthetic data. Open for others to join, contribute, and utilize—building a shared resource ecosystem.

VI/ IMU drift and calibration inconsistencies are a big challenge, also noisy data and orchestrating long tail environments and data collectors. There is no magic pill here. In post-processing, applying sensor fusion techniques (filters, wavelets etc.) to achieve sub-degree precision in processed data, capturing natural variations in how people perform tasks. A long way to go. But each collected hour of data will worth it.

V/ In post-processing, our eidon-sim pipeline handles quaternion-to-Euler conversion, forward/inverse kinematics, and filtering. We aim to collect and process 1000s of hours of data from diverse tasks like grasping and assembly, turning raw inputs into clean 3D trajectories.

IV/ Eidon SYM acts as the hub: It aggregates IMU data and syncs it with timestamped video from a head (or chest)-mounted phone camera. This creates multimodal datasets—skeletal poses paired with egocentric visuals—for training Vision-Language-Action models.

III/ Hardware details: we have created custom PCBs with 9-DOF IMUs for fused orientations. We've focused on secure mounting to minimize wear variability and ensure consistent calibration, reducing noise in real-world collections.