ABC-130K: The largest open source teleoperation dataset

The momentum behind robotics is unprecedented. New advancements up and down the stack have driven incredible energy and interest into the space. They have provided the world with a tangible look at the path towards general-purpose robots. It is much needed: as language models have created fully agent-driven experiences, we are left painfully aware of the gap between what is possible in the digital versus the physical. The researchers' scramble for data is one we know well from our days at UC Berkeley and beyond. Too often, researchers are forced to piece together whatever bits of data they can find from existing datasets like DROID or Open X-Embodiment. All this effort is just for data collection, let alone processing, language annotations and quality control. We think the robotics world deserves more, and better. So, we teamed up with the authors behind ABC: Scalable Behavior Cloning with Open Data, Training, and Evaluation to do something about it. More data, higher quality, broader access As part of the broader paper, we are excited to release the following datasets for research purposes. ABC-130K contains over 130,000 episodes across 195 bimanual manipulation tasks. ABC-Sim contains teleoperation data in simulation for 10 tasks and is coming soon. ABC-Eval contains real evaluation trials of various checkpoints and hyperparameters for 3 tasks and is coming soon. All data are on bimanual station setups using YAM arms from I2RT. All datasets are provided under the Apache 2.0 license. ABC-130K is the largest open source teleoperation dataset to-date. It combines data scale while maintaining high quality, all while using the accessible YAM setup that has gained popularity in the last couple of years. Example task categories include pick and place, folding, insertion and ejection, tool use, assembly and disassembly, tying and untying, and more. We also made XDOF's evaluations service available to the authors. Establishing a foundation for behavior cloning Thank you to UC Berkeley lead authors on this work and collaborators for making this all possible. The data is just one part of the foundation laid out by this work. Our hope is to provide the infra for robot learning that we wish we had.