The Soft Robotics Lab at ETH Zurich is open-sourcing the ORCA Hand, a 17-degree-of-freedom anthropomorphic robotic hand with integrated tactile sensors that one person can assemble in under eight hours. All design files, control code, simulation environments, and step-by-step assembly guides are available on GitHub and the ORCA dashboard.
The Hardware Gap in Dexterous Manipulation
Dexterous robotic hands that approach human capability have historically been expensive, fragile, and proprietary. Reinforcement learning and imitation learning continue to make massive strides on the software side, but labs often cannot deploy their policies on real hardware without spending tens of thousands on systems like the Allegro Hand.
ORCA is changing that equation. A research group can go from nothing to a fully functional hand within a day using 3D-printed structural parts and off-the-shelf components. The hand mirrors human proportions with an opposable thumb, MCP, PIP, and ABD joints, plus an actuated wrist, which means it works directly with human hand interaction datasets and simplifies teleoperation retargeting.
Popping Joints and Tendon Routing
The ORCA v1 runs 16 finger DOFs and one wrist DOF on a tendon-based system powered by Dynamixel servos. Tendons are routed through the center of rotation of each joint, which keeps friction low and performance consistent over time.
The standout design detail: joints that safely dislocate under excess load instead of snapping. If something pops, you push it back in. No replacement parts, no downtime. Combined with auto-calibration and a built-in tensioning system, this makes ORCA one of the more maintenance-friendly tendon-driven hands out there. Integrated tactile sensors sit under a soft silicone skin, and teleoperation works out of the box with ROKOKO Gloves and Apple Vision Pro.
10,000+ Cycles, No Failure
The team has run continuous grasping for 2,250 cycles over 2.5 hours with zero tendon slack, zero motor shutdown, and zero breakage. They stopped because the point was proven, not because something failed.
In a separate imitation learning deployment, a trained pick-and-place policy ran continuously for over 7 hours straight, roughly 2,000 grasps, with no human intervention on the hardware. The paper reports the hand surviving more than 10,000 continuous cycles, about 20 hours of operation, without hardware failure.
Sim-to-Real in One Hour
Using IsaacGym with 4,096 parallel ORCA models and domain randomization, the team trains RL policies for in-hand ball reorientation. After one hour of training, a policy transfers directly to the physical hand with zero fine-tuning, successfully reorienting a tennis ball along a target axis.
Open Everything
The project lives across several repos under github.com/orcahand: orca_core for Python control, orca_sim for simulation, orcahand_description for MJCF/URDF models, and orca_retargeter for teleoperation. MIT and Creative Commons licensed. The shop sells assembled hands and kits for those who prefer not to self-source.
The ORCA Hand is being developed at ETH Zurich's Soft Robotics Lab by Clemens C. Christoph, Maximilian Eberlein, Filippos Katsimalis, Arturo Roberti, Aristotelis Sympetheros, Michel R. Vogt, Davide Liconti, Chenyu Yang, Barnabas Gavin Cangan, Ronan J. Hinchet, and Robert K. Katzschmann.
Source: ORCA Hand | ETH Zurich SRL | GitHub | arXiv