ORCA: An Open-Source, Reliable, Cost-Effective, Anthropomorphic Robotic Hand for Uninterrupted Dexterous Task Learning

TL;DR

ORCA (ETH Zurich Soft Robotics Lab, IROS 2025) is an open-source 17-DoF tendon-driven anthropomorphic robotic hand with integrated fingertip tactile sensors, 3D-printable, assemblable in <8 hours by one person for under 2,000 CHF in material cost. Design innovations — poppable pin joints that dislocate instead of breaking, tendons routed through joint centers of rotation to enable model-free auto-calibration, and a belt-driven actuated wrist — target the “hardware is the real bottleneck for dexterous manipulation” thesis directly. Demonstrated with 2,250-cycle reliability testing, sub-0.2s latency comparable to the direct-driven LEAP hand, zero-shot sim-to-real in-hand ball reorientation via A2C, and a 7h 17min continuous imitation-learning pick-and-place with no hardware intervention.

Key claims

17-DoF tendon-driven design with five fingers (index-to-pinky each 3 DoF: MCP, PIP, ABD), four-DoF opposable thumb (IP, MCP, ABD, CMC), and one wrist DoF assembled in <8 hours for <2,000 CHF material cost [§II, §IV-D].
Poppable pin joints — bearings held in arc-shaped grooves that dislocate rather than break under excessive load — combine axial stability of pinhole joints with robustness of rolling-contact joints while remaining trivial to reassemble [§II-B].
Auto-calibration is enabled by routing each tendon through or near the joint’s center of rotation, which makes joint-to-motor mapping approximately linear so the transmission ratio can be recovered by moving each joint to its mechanical stops and interpolating, without proprioceptive sensors [§II-F].
2,250 grasping cycles of a plush object over 2.5 hours of continuous operation with no hardware failure and constant per-cycle peak motor current, terminated by choice rather than by breakage [§III-A, Fig. 4-C].
On a sine-wave joint-tracking benchmark against the direct-driven LEAP hand, ORCA achieves comparable joint-angle accuracy and average end-to-end latency <0.2s (most of it software) at 2–5 Hz [§III-B, Fig. 4-A].
Zero-shot sim-to-real RL: A2C-trained ORCA policy for in-hand ball reorientation, trained in 1 hour with 4,096 parallel envs in IsaacGymEnvs and domain randomization, transfers to the physical hand [§III-C].
Diffusion-transformer imitation-learning policy trained on 214 mocap-glove demos runs continuously for 7h 17min (~2,000 grasp cycles) on the pick-and-place task with no hardware intervention; masked-cube image inputs reach 87% success across six spatial sub-areas vs 50% for raw-RGB baseline [§III-D, §IV-B, Fig. 7].
Integrated fingertip FSR tactile sensors register forces down to ~0.7 N with 100% accuracy over 10 cycles despite a 1 N rated minimum trigger, but degrade after ~2,000–4,000 grasp cycles as the silicone skin wears and ~4,500–7,000 cycles as the copper signal wires snap at the MCP/ABD joint boundary [§III-E, §IV-C].

Method

Each non-wrist joint is driven by two antagonistic tendons (0.4 mm braided Nylon), one flexor and one extensor, routed through smooth metal pins and PTFE tubes to keep them off the PLA structure and manually retensioned via a ratchet spool that locks in one direction. Bearings sit in arc-shaped grooves that release under overload — a “poppable” pin joint that combines the kinematic simplicity of a pinhole joint with the impact tolerance of a rolling-contact joint but reseats in seconds. The DIP joints (fingers 2–5) are intentionally omitted to simplify assembly and free room for fingertip tactile sensors, on the argument that DIP is not directly actuated in the human hand anyway. The wrist uses a GT2 fiberglass-reinforced timing belt rather than tendons to handle higher torque loads and avoid slack buildup. Fingertip tactile is FSR (RP-C7.6-ST) on FDM-printed backplates under silicone skin, wired through internal PTFE tubes to an Arduino Nano Every voltage divider, providing binary contact per fingertip. Model-free auto-calibration exploits the near-linear motor-to-joint map enabled by CoR-routed tendons: joints are moved to their CAD-known mechanical limits and the transmission ratio is computed by linear interpolation, avoiding the cost of encoders. Reinforcement learning uses A2C in IsaacGymEnvs with domain randomization; imitation learning uses a diffusion-transformer policy on top of a Franka Emika Panda arm, with demonstrations collected via a motion-capture glove and retargeted to hand joint space by energy-based minimization.

Results

Reliability: 2,250 grasp cycles + wrist flexion in 2.5h of continuous operation, no failure, constant peak motor current per cycle [§III-A, Fig. 4-C].
Accuracy/latency: joint-angle tracking on 2 Hz and 5 Hz sine waves matches direct-driven LEAP hand; average latency <0.2s [§III-B, Fig. 4-A].
RL: zero-shot sim-to-real ball reorientation after 1h A2C training in sim [§III-C, Fig. 6-B].
IL: 7h 17min continuous pick-and-place autonomy, ~2,000 grasp cycles, no hardware intervention [§IV-B, Fig. 1-B].
IL policy ablation (60 iterations across 6 sub-areas): masked-cube 87% success, hybrid 70%, raw-RGB 50% [§IV-B, Fig. 7].
Tactile: 0.7 N minimum detectable force at 100% over 10 cycles vs 1 N rated minimum, degradation to 1.7 N by 4,000 cycles from silicone wear, copper wires snap at MCP boundary by 4,500–7,000 cycles [§III-E, §IV-C].
Cost: <2,000 CHF material, <8h single-person assembly [§Abstract, §IV-D].

Why it’s interesting

ORCA and Aero Hand Open: low-cost, open-source, lightweight, anthropomorphic dexterous hand (Aero Hand Open, $314 / 16 joints / 7 motors) stake out the two current open-source-anthropomorphic-hand price points and hit opposite ends of the DoF/cost/actuation-count trilemma — ORCA keeps one motor per non-wrist joint (17 motors, 17 DoF, <2,000 CHF) and integrates fingertip tactile, while Aero Hand Open shares 7 motors across 16 joints via tendon coupling to hit$ 314 and skips fingertip tactile entirely. Both agree on tendon-driven + backdrivable as the right substrate for dexterous manipulation, and both use MuJoCo/IsaacGym-trained zero-shot sim2real as the validation demo. Against Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1) mimic hand M1, ORCA sits at the open-source-research-hand end while mimic sits at the vertically-integrated commercial stack end — mimic gets whole-hand implicit force sensing via motor-current backdrivability across every joint whereas ORCA integrates dedicated FSR tactile only at fingertips. On the reliability-testing axis, ORCA’s 2,250-cycle self-terminated test sits alongside RLWRLD’s How we actually test a robot hand All Hands Up! vendor-audit archive as one of the few filed hardware pages that publishes stress-test numbers; ORCA’s Kapandji equivalent isn’t stated but the 17-DoF joint layout suggests a strong opposition score.