Eidon AI shuts down and releases 9TB of robot-training video

Why it matters

Eidon AI's shutdown converts an expensive private data-collection effort into open robotics infrastructure, while exposing the imbalance and privacy tradeoffs inside real-world training corpora.

Eidon AI has shut down and released 9.05 TB of first-person household video for training robotics models, turning the Austin startup's principal data asset into an open dataset after two years of operation.

https://x.com/LeRobotHF/status/2102011968026554652

The release surfaced Sunday through Hugging Face's LeRobot account, which said Eidon AI was winding down. The underlying Eidon Tracker POV dataset contains 13,451 recordings totaling 1,273.8 hours, captured by 27 contributors performing chores in their homes. Eidon AI published the files under a CC-BY-4.0 license, permitting commercial use with attribution.

Eidon AI was founded in 2024 by Peter Toth (@peter_van_toth) and Sam Padilla (@theSamPadilla). Toth previously worked as an AI researcher at DeepMind. Padilla, a computer science graduate of Morningside University, had been an engineer and product manager on Google Cloud's Web3 team before leaving Google to build Eidon AI.

The founders initially framed Eidon AI as a decentralized network that would pay people to contribute data for AI training. Padilla wrote that contributors should be able to own, shape and benefit from AI systems, while Eidon AI's early documentation described plans to record contributions through blockchain infrastructure. In November 2024, Eidon AI raised a $3.5 million seed round led by Framework Ventures, with cyber.Fund participating. Padilla said at the time that the financing would help Eidon AI build its team and begin with data collection.

That data collection operation produced something considerably more concrete than the planned network. Contributors wore a head-mounted camera and a seven-sensor inertial measurement unit harness while folding laundry, cleaning, washing dishes and cooking. The sensors recorded orientation at 24 Hz across both hands, both forearms, both shoulders and the chest, producing about 779 million rows of motion data.

The release separates the corpus into 9.05 TB of MP4 video and 9.5 GB of sensor streams, joined by recording IDs. Eidon AI also published a separate 1.55 TB bucket containing 306 hours of video without corresponding sensor data.

The paired footage gives robotics researchers a view of human manipulation from the perspective a wearable camera or robot might see, while the sensor harness supplies a matching record of upper-body movement. That combination can support work on imitation learning, activity recognition and converting human demonstrations into robot actions. Researchers can load the metadata and sensor tables directly through Hugging Face, although downloading the full video archive requires substantial storage and bandwidth.

The dataset carries material limits

The headline duration overstates how varied the corpus is. Folding laundry accounts for 860.2 hours, or 67.5% of the paired dataset. Cleaning contributes another 276.3 hours. Washing dishes, cooking, drawing, knitting, making beds, watering plants and organizing collectively make up the remaining 10.8%.

The contributor pool is also concentrated. Eidon AI's documentation says the five largest contributors produced 57.2% of all recorded hours. Researchers are advised to divide training and evaluation sets by contributor rather than randomly, since a random split could place footage from the same person, home and camera equipment on both sides of an experiment.

The IMU data requires similar care. Orientation measurements are present across the corpus, but raw accelerometer, gyroscope and magnetometer readings appear in only 2,841 recordings, or 21.1% of the total. Another 129 recordings have fewer than all seven sensor positions.

Eidon AI ran automated quality checks for hand visibility, lighting, blur and camera stability. The release labels 11,841 recordings representing 1,160 hours as valid. It also retains 1,138 flagged recordings and 472 invalid recordings, allowing researchers to filter them out or use them as negative examples.

The footage was recorded inside private homes and remains unprocessed. Eidon AI says adult contributors provided written consent covering research, commercial use and public redistribution. Metadata identifiers and device addresses were removed, but the videos were not blurred or redacted. Faces, documents, computer screens and home interiors may appear incidentally, according to the dataset card.

Eidon AI's final act preserves the part of its business that robotics developers can use immediately. The decentralized network and its planned contribution economy have ended with Eidon AI. The human demonstrations, motion traces and collection hardware's output will remain available to anyone willing to download them.