ORCA: A Platform for Open-Source Dexterity Research

A new open-source platform called ORCA aims to make dexterous robot learning more accessible by unifying the fragmented software tools researchers currently rely on, according to a paper submitted to arXiv in June 2026 [1]. The work, led by Francesco Capuano, addresses a persistent divide in robotics manipulation research. Most learning-based studies use two-finger parallel grippers because they are effective, affordable, and easy to teleoperate, but their simple form factor limits what a robot can do with a single hand [1]. Anthropomorphic hands offer a more natural path toward human-like dexterity and can learn from human video, yet the software needed for control, simulation, teleoperation, and retargeting has remained scattered across incompatible code bases, disconnected from the broader robot-learning ecosystem [1]. The ORCA stack consolidates low-level control, simulation, teleoperation from consumer devices, and hand retargeting behind a single interface [1]. It also integrates natively with LEROBOT, a popular robot-learning framework, so researchers working with dexterous hands can use the same data, training, and evaluation pipelines already established for simpler gripper-based systems [1]. The team demonstrated the full workflow by collecting expert demonstrations of an in-hand reorientation task via a consumer-grade VR headset, training an autonomous policy with LEROBOT, and evaluating it in a reproducible setup [1]. The paper was posted on arXiv, the open-access e-print repository that has hosted more than two million articles since its launch in 1991 and currently receives roughly 24,000 submissions per month [8]. The ORCA manuscript, dated June 12, 2026, spans 24,650 KB and was submitted under the Computer Science and Robotics category [1]. By open-sourcing the entire stack, the authors aim to provide a shared foundation for dexterous-manipulation research, lowering the barrier for labs that want to move beyond parallel grippers without having to stitch together one-off software solutions [1]. The platform’s reliance on consumer-grade teleoperation hardware, such as VR headsets, further reduces the cost and complexity of generating demonstration data for learning-based policies [1].