Skip to content

Aero Hand Open: low-cost, open-source, lightweight, anthropomorphic dexterous hand

TetherIA releases Aero Hand Open, an open-source anthropomorphic robotic hand costing 314inparts,weighingunder400g,with16jointsdrivenby7motorsviaatendondrivenarchitecture.Thedesignisfullybackdrivable,supportspositionandtorquecontrolplusinlinetendonforcesensing,holds 10Nperfingertipand 18kgtotalpayload,andisvalidatedbya1millioncyclefullspeedreliabilitytestandanactuationlevelMuJoCorealzeroshotsim2realRLpolicy.Thebetisthathandhardwaredoesnthavetobe314 in parts, weighing under 400 g, with 16 joints driven by 7 motors via a tendon-driven architecture. The design is fully backdrivable, supports position and torque control plus inline tendon-force sensing, holds ~10 N per fingertip and ~18 kg total payload, and is validated by a 1-million-cycle full-speed reliability test and an actuation-level MuJoCo → real zero-shot sim2real RL policy. The bet is that hand hardware doesn't have to be 10k+ proprietary to be research-grade — a $314 open BOM plus tendon sharing can hit the specs that dexterous-manipulation policies need.

  • 7 motors drive 16 joints via low-friction shared tendon routing, keeping mass under 400 g while recreating natural hand synergies [§Tendon-Driven Architecture].
  • The thumb has 3 active DoFs supporting opposition, pinch, and in-hand reorientation [§Key Characteristics].
  • Cable routing plus low gearing make the hand 100% backdrivable, enabling safe interaction and torque control [§Key Characteristics].
  • Active load torque is 10 N per fingertip, sufficient to pull a drill trigger; the whole hand can lift a ~18 kg water jug [§Hand Payload].
  • Reliability test: 1 million cycles at full speed over 5 × 24 hours without failure [§Reliability Test].
  • Actuation-level RL policy trained in MuJoCo transfers zero-shot to the real hand [§RL policy].
  • Bill of materials totals $314; full mechanical design, firmware, and MuJoCo assets are open-sourced at github.com/TetherIA/aero-hand-open [§Abstract, BibTeX].

The core mechanical bet is tendon sharing: rather than one motor per joint (which drives cost, mass, and inertia up), each of the 7 motors distributes force through low-friction cable pathways to multiple joints, coordinated to reproduce natural hand synergies across the 16 joints. Compliant cable runs absorb impact energy, which both cushions object interaction and simplifies contact-rich control. The hand exposes three control modes — position, torque, and inline tendon-force sensing — the last of which gives a cheap proxy for contact and load without adding fingertip tactile arrays. The RL demonstration trains an actuation-level policy directly in MuJoCo (not joint-level, so the tendon coupling is respected in simulation) and deploys it zero-shot to hardware, sidestepping the joint-space sim2real gap that has historically been the failure mode for tendon-driven hands.

  • Cost: $314 BOM.
  • Mass: <400 g.
  • DoF: 16 joints, 7 motors, 3 active thumb DoFs.
  • Payload: 10 N per fingertip active load torque; ~18 kg total lift (5-gallon water jug).
  • Reliability: 1M cycles at full speed, 5 × 24 h.
  • Speed: catches a flying tennis ball; webcam-based teleop runs on a laptop.
  • Sim2real: MuJoCo-trained actuation-level policy transfers zero-shot.

This slots directly against Solving Dexterity: A Full-Stack Approach (mimic hand M1 + wearable U1) — mimic hand M1 is also tendon-driven and highly backdrivable, but sits at the vertically-integrated commercial stack end (15 active DoF, dedicated tactile array, custom wearable data-collection glove) while Aero Hand Open sits at the cheap-open-hardware-substrate end ($314 BOM, no fingertip tactile, tendon-force as contact proxy). Both agree on tendon-driven + backdrivable as the right actuation choice for dexterous manipulation — the disagreement is on which layer of the stack should be open. It also complements Tactile Genesis: Exploring Tactile Sensors at Scale for Learning Dexterous Tasks‘s finding that tactile-sensor placement dominates type: Aero Hand Open ships without fingertip tactile at all and uses tendon-force as the contact signal, which is the sparsest end of that design axis. Finally, the actuation-level MuJoCo sim2real recipe is a data point for the sim2real-for-locomotion / BC-for-manipulation dichotomy Danfei Xu named in Danfei Xu — two root-level paradigm shifts in robot learning: Sim2Real for locomotion, Behavior Cloning for manipulation — Aero Hand Open runs a manipulation policy through the sim2real pipeline, not BC, which is unusual for a five-finger hand.