Source-linked AI summary
Aero Hand Open: A Simulation-Ready Tendon-Driven Hand for Dexterous Manipulation Learning
Nan Wang, Mohit Yadav, Jonathan Wulff, Aidan Rosenbaum, Kezhou Chen, Yuvan Sharma, Xu Dong, Yiwei Tao
TL;DR
Tendon-driven hands offer affordable anthropomorphic dexterity but are difficult to simulate and learn because shared cables create underactuated, coupled transmissions. Aero Hand Open combines a transmission-faithful simulation, identified actuation map, and hardware-constrained reinforcement learning package; a cube-rotation policy trained entirely in simulation transfers without fine-tuning or state estimation, with documented mapping limitations.
Problem
Tendon-driven hands are affordable and anthropomorphic, but their underactuated cable transmissions are difficult to represent in simulation and their coupled joints are not independently commandable.
Method
The paper releases a tendon-level MuJoCo model, a bidirectional identified actuation map including thumb coupling, and a reinforcement learning package restricted to hardware-available signals.
Results
A cube-rotation policy trained entirely in the model transfers to the hand without fine-tuning or state estimation; model residuals are 0.29–0.45 mm RMS under shared command streams.
Takeaways & Limitations
The released pipeline makes zero-shot deployment of policies trained in simulation possible for this tendon-driven hand.
Takeaways & Limitations
The simulator does not reproduce thumb abduction-to-flexion coupling, so the actuation map compensates for it in software.
Abstract
from arXiv · showhide
Tendon-driven hands are anthropomorphic, and moving the actuators off the joints is what makes a hand of this capability affordable to build. Two effects produce that saving. Routing force through a cable removes the requirement that a motor fit inside the joint it drives, so smaller and cheaper motors suffice, and one motor can drive several joints through a single cable, so fewer motors are needed. They are also harder to learn on than a direct-drive hand. The underactuated transmission that produces the saving is itself difficult to represent in a simulator, and the joints one cable drives are not independently commandable. We present Aero Hand Open, a tendon-driven anthropomorphic hand that is released simulation-ready. Three things ship with it. A simulation model reproduces the cable transmission itself. An identified actuation map connects that model to the motor commands in both directions, including the three-way coupling of the thumb. A reinforcement learning package trains policies for the hand. Together they let a policy be trained entirely in simulation and run on the hand with no fine-tuning and no state estimation. We release the mechanical design, the simulation model, the identified mapping, the training environment and the deployment stack.
1 Introduction
Aero Hand Open addresses the cost and learnability trade-off of tendon-driven dexterous hands by releasing an affordable hand together with a transmission-faithful simulation and sim-to-real learning pipeline.
- Motivation: Dexterous manipulation requires multi-finger hands for irregular objects, in-hand rotation, and thumb opposition, but such hands have been expensive or difficult to learn on.Tendon-driven transmissions improve affordability through remote and shared actuation, while their underactuation complicates simulation and control.
- Platform: Aero Hand Open is a five-finger, sixteen-joint hand with seven motors, 374 g mass, and a $314 bill of materials.Its design files, firmware, and control stack are released, and one hardware configuration covers all 33 GRASP taxonomy grasp types.
- Simulation model: The MuJoCo model represents every actuation cable, coupling cable, and return spring using routing geometries taken from the CAD model.This models the tendon transmission itself rather than treating joints as independent position actuators.
- Sim-to-real interface: The identified actuation map connects simulation and motor commands in both directions while representing the thumb’s three-way coupling.The map is validated channel by channel kinematically and dynamically, and domain randomisation supports zero-shot transfer.
- Learning pipeline: The released reinforcement learning package provides the training environment, reward and randomisation settings, and deployment node using hardware-available observations and actions.The pipeline is demonstrated on in-hand cube rotation.
2 Hand Design
Aero Hand Open combines an underactuated, repairable mechanical design with tendon routing, minimal sensing, compliant contact, and broad grasp capability. Its durability and repeatability measurements characterize the practical operating envelope of this low-cost hand.
- Design overview: The hand has five fingers, sixteen revolute joints, seven motors, a 374 g mass, and dimensions of 198 × 95 × 53.5 mm.Fifteen joints are cable-driven, while thumb abduction is driven directly through a short linkage.
- Finger transmission: Each underactuated finger uses one cable to drive MCP, PIP, and DIP flexion, with passive extension and staged closure set by spring stiffness.The MCP closes first; the distal joints curl after the MCP reaches its 1.57 rad limit.
- Thumb transmission: The thumb uses two cables and direct CMC abduction to control three mutually coupled flexion channels across a 0–100° abduction range.CMC flexion, MCP, and IP motion share routing, producing three-way coupling and thumb opposition.
- Sensing and control: The only proprioception is provided by seven 16-bit motor encoders, defining the hardware-available observation space for learning.This prevents training on joint states unavailable during deployment.
- Grasp capability: The hand spans the full range of human grasping with one hardware configuration, while each finger and the thumb exert approximately 12 N at the tip.The hand completes a full open-and-close cycle at about 1.2 Hz.
- Durability: More than 400,000 full-actuation cycles occur without cable failure or loss of range, corresponding to over 90 h at the nominal 1.2 Hz rate.The reported endurance exceeds the actuation accumulated during a long training-and-deployment campaign.
- Durability: The cable has about 680 N tensile strength against approximately 66 N worst-case actuator tension, providing roughly a tenfold margin.The servo’s 0.59 N m stall torque acts on a 9.0 mm spool to set the actuator-tension bound.
3 Tendon-Driven Simulation Model
The MuJoCo model represents Aero Hand Open’s 16-joint, seven-actuator transmission with spatial tendons, wrapping geometries, springs, and explicit joint couplings. Its controls and sensors mirror the seven motor spools and available encoder measurements, while simulation choices support stable parallel training.
- The released MuJoCo model contains 16 hinge joints, 20 spatial tendons, and 7 actuators.
- Each finger’s actuation, coupling, and extension elements are represented as spatial tendons, with specified spring stiffnesses and rest lengths.Finger cables drive the distal phalanx through wrapping cylinders, while coupling cables link PIP and DIP joints.
- Wrapping cylinders taken from CAD reproduce physical pulley routing, with explicit wrap-side specifications for the tendon paths.The routing geometry is used to model the physical transmission rather than approximate it with independent joint actuators.
- Joint equality constraints additionally enforce PIP–DIP coupling in the fingers and MCP–IP coupling in the thumb.
- Seven control inputs correspond one-to-one with the motor spools, while six tendon-position sensors and one joint-position sensor expose encoder-level quantities.The thumb CMC abduction channel is modeled with a direct joint actuator, matching its short-linkage hardware routing.
- The simulator uses semi-implicit Euler at a 10 ms timestep and manually curated primitive collision geometries and contact pairs.Implicit joint-damping integration is disabled because earlier configurations produced non-finite accelerations after tens of seconds.
4 Actuation Mapping and System Identification
The actuation map represents cable travel, joint motion, and motor commands for the underactuated fingers and coupled thumb. It combines CAD-derived winding coefficients with identified bidirectional mappings, while validation exposes residual coupling and slope limitations.
- Joint-to-actuation model: The linear actuation model uses effective winding radii in millimetres of cable travel per radian, with sign encoding cable take-up or release.Every finger coefficient is positive; the thumb map contains one negative entry for cable release during CMC flexion.
- Joint-to-actuation model: Winding coefficients are computed from CAD by comparing routed cable paths at the two endpoints of each joint’s travel.The construction evaluates cable take-up over the joint’s angular range, with unaffected tendon length cancelling from the difference.
- Actuation map: Each finger’s single cable drives MCP, PIP and DIP joints together, whereas the thumb requires a three-channel map with cross terms between abduction, CMC flexion, MCP and IP.The thumb flexor relation includes 2.50 q_abd − 2.50 q_cmc + 9.44 q_mcp + 12.50 q_ip.
- Actuation map: The simulator and hardware exchange commands through separate forward and inverse interfaces, using affine maps for fingers and an identified map for the thumb.The thumb mapping is applied in both directions between simulator lengths, motor commands and encoder-derived observations.
- Validation: Endpoint anchoring makes the command map exact at interval extremes, but its interior slope is inherited rather than directly measured.Measured, coefficient-derived and simulated finger excursions were 44.45 mm, 45.26 mm and 45.72–46.21 mm, respectively.
- Validation: The kinematic model reproduced finger cable excursions within 0.04, 0.32, 1.40 and 0.41 mm, but thumb channels differed by 32.7% and 31.2%.The remaining thumb discrepancy is associated with a flexor effective-radius shortfall after software restores the missing abduction coupling.
- Validation: 0.29–0.45 mm RMS finger tracking residuals and 0.63 mm RMS for the thumb flexor were obtained in settled dynamic validation.Simulation rise times were consistently 10–40 ms faster than hardware, and the largest amplitude mismatch occurred when sweeping CMC flexion against the thumb flexor.
5 Learning In-Hand Cube Rotation
The paper trains an in-hand cube-rotation policy in MuJoCo using hardware-available observations and actions, domain randomisation, and a reward combining rotation, grasp stability, and command smoothness.
- Task and training: PPO trains the policy in MuJoCo Playground to rotate a 50 mm, 69.2 g cube continuously while maintaining its grasp.Physics runs at 100 Hz and the policy at 20 Hz; episodes last 500 control steps, or 25 s.
- Observation and action: The policy observes seven measured channel values, the previous action, and independently sampled sensor noise restricted to hardware-available signals.Cable channels receive 5 mm noise and the abduction channel receives 0.05 rad noise; the critic additionally receives an 81-dimensional privileged state.
- Observation and action: Actions are seven-dimensional target offsets from a fixed nominal grasp, with per-channel scales chosen to remain inside calibrated actuation ranges.Thumb channels use smaller scales because their ranges are three to five times narrower than a finger’s.
- Reward: The reward combines vertical-axis rotation, a fall penalty, and action-change regularisation that limits cable travel per 50 ms control period.Cube linear velocity, pose regularisation, torque, and energy terms are weighted zero.
- Randomisation and optimisation: Training uses independently resampled physical parameters, while the reported policies retain fixed cube friction of 0.3 and mass of 69.2 g.The policy is trained for 3 × 10^8 environment steps with 8192 parallel environments.
6 Results
The policy rotates the cube continuously in simulation and transfers to the physical hand without fine-tuning. Hardware execution sustains rotation, while replay analysis localises the remaining sim-to-real error to the thumb.
- Simulation: The simulated policy holds the cube for 15 s, turns it through 7.64 rad, and achieves a mean rate of 0.51 rad/s.This corresponds to 1.22 revolutions in the plain-MuJoCo deployment replica.
- Zero-shot deployment: The physical deployment uses the same 20 Hz policy and seven actuator states, mapping encoder readings into simulation units without fine-tuning.The deployed node reconstructs thumb channels by solving the inverse actuation map.
- Zero-shot deployment: The hardware rotates the cube through a full revolution in about 55 s and sustains that rotation stably.Cube pose is not logged on the hardware, so its rotation rate is not measured there.
- Zero-shot deployment: Deployment differs from training through 20–30% larger cable scales, a nominal pose closed by 4–8 mm, and absent observation noise.Both systems run at 20 Hz and retain the same thumb scales and nominal values.
- Error analysis: Open-loop replay agrees within a millimetre on the finger channels and localises the residual sim-to-real gap to the thumb and policy feedback.The replay comparison separates transmission-model error from feedback-related error.
7 Conclusion and Outlook
Aero Hand Open combines an open, lightweight tendon-driven hand with a transmission-faithful simulation and hardware-matched learning pipeline. The reported model supports zero-shot transfer of an in-hand rotation policy, while broader task coverage remains future work.
- Conclusion: The hand weighs 374 g, is fully 3D printed and open-sourced, and spans the full range of human grasping.Its released mechanical design, firmware, and control stack support repair by reprinting worn or broken parts.
- Conclusion: The pipeline represents the cable transmission, identifies bidirectional actuation maps, and restricts policy observations and actions to hardware-provided signals.This addresses the difficulty of learning tendon-driven hands in conventional joint-actuated simulation models.
- Results: Finger cable excursion errors are 0.04–1.40 mm, dynamic residuals are 0.29–0.45 mm RMS, and an in-hand cube-rotation policy transfers without fine-tuning.The routing geometry is modelled rather than fitted through moment arms, and deployment uses no state estimation in the loop.
- Outlook: The released task suite currently needs extension beyond in-hand rotation to exercise grasps across the full range the hand can reach.The hardware, hand model, simulation environment, and analysis scripts are open-sourced.
8 Resources
The project provides an open-source repository and online documentation for Aero Hand Open.
- The open-source repository is available at github.com/TetherIA/aero-hand-open.
- Online documentation is available at docs.tetheria.ai.