The team behind the ABC project provided 3,500 hours of bimanual robot data and hardware specifications for academic research. This open stack encompasses training code and simulation tools used to collect 130,000 episodes across 195 tasks using a setup costing $8,000. The release further includes 400 hours of simulation data for 24 tasks and 5,850 labeled policy evaluation episodes.

The ABC research was accepted for presentation at the CoRL 2026 conference. Baseline science provided by the authors includes a 0.91 sim-to-real correlation on task progress and analysis of compute scaling laws regarding training steps and batch size. The team also evaluated which offline metrics are most effective for conditioning robot policies to provide a foundation for future academic study.