Orca: Open-Source Robotic Hand for Uninterrupted Dexterous Task Learning (orca.ethz.ch)

Researchers released ORCA v1, an open-source, anthropomorphic robotic hand designed to make dexterous manipulation hardware affordable, reliable, and easy to reproduce. ORCA is a 17-DoF tendon-driven platform (16 finger DOF + 1 wrist) with integrated tactile sensing, an opposable thumb, and human-like MCP/PIP/ABD joints. It is largely 3D-printable, has a bill-of-materials under ~2,000 CHF, can be assembled in under 8 hours, and is distributed with STLs, control code, step-by-step assembly/repair guides and permissive MIT/Creative Commons licenses. Mechanical design choices—“poppable” joints that dislocate safely, low-friction tendon routing through joint centers, auto-calibration, modular layout and tensioning systems—prioritize robustness and easy repair; the hand sustained thousands of cycles in continuous tests (2,000+ pick-and-place cycles over 7+ hours, 2,250 grasps over 2.5 hours and claims of >10k cycles durability). For the AI/ML community ORCA lowers the hardware barrier for long-duration learning, sim-to-real transfer, and benchmarking. The authors demonstrated uninterrupted imitation-learning deployment (7+ hours) and zero-shot sim-to-real RL: training 4,096 simulated ORCA hands in Isaac Gym with an advantage actor-critic agent and domain randomization for one hour produced a policy that reoriented a tennis ball on the real hand. By providing a low-cost, standardized, repairable platform with built-in sensing and human-like kinematics, ORCA enables reproducible dataset collection, teleoperation/retargeting, and faster iteration of RL and imitation algorithms in real-world dexterous manipulation.