Dexterity Contact

Dexterity Card Protocol

How Robotics Center measures a robot hand. Published as a draft so the method is public, and open to comment, before any card is issued.

v0.1 · Draft for comment Published 2026-09-27
Comment on the draft

Scope

A Dexterity Card records how one physical hand behaves on our bench, bound to its serial number. Six measurements, one fixed procedure, the same for every brand.

This is v0.1, a draft. Target poses, trial counts, loads and durations below are our starting proposal and may change after review. No card has been issued under this protocol, and this page contains no measured values.

The six measurements

Instruments: a force gauge, a high-speed camera and ArUco fiducial markers. Commands and joint readback go through the vendor's own SDK, in its default control mode. Measurements run in the order listed.

1 · Fingertip repeatability mm
  • Definition. How far the fingertip lands from where it landed before, when the same joint command is sent repeatedly. Computed per finger as the mean distance from the centroid of reached positions plus three standard deviations (adapted from ISO 9283 pose repeatability).
  • Unit. mm.
  • Setup. Hand rigidly mounted (see Conditions). One ArUco marker on each fingertip and a reference marker on the fixture. Camera intrinsics calibrated with a ChArUco board at the start of the session. Marker mass recorded.
  • Trials. 3 target poses × 30 cycles, per finger.
  • Reported. Per finger, per pose: repeatability, mean ± sd distance, max distance. The card shows the worst finger.

Procedure

  1. Define three joint-space targets: open, half-closed, tip pinch. Record the exact joint values.
  2. Approach each target alternately from fully open and fully closed, so direction-dependent error is included.
  3. Wait 1 s after the SDK reports the move complete, then average 30 camera frames for the fingertip pose relative to the reference marker.
  4. Repeat for 30 cycles per target. Log SDK joint readback at each capture.
  5. Compute the centroid and distances per finger and pose.

Known limitations: single-camera marker pose is less accurate along the viewing axis than across it, so we state the measured resolution next to every value and report anything below it as "< resolution". Unloaded fingers only; repeatability while holding an object is not measured in v0.1.

2 · Command-to-motion latency ms
  • Definition. Time from the SDK command call on the host to the first visible fingertip motion.
  • Unit. ms.
  • Setup. High-speed camera at ≥ 1000 fps. A host-driven LED in frame switches on in the same call that sends the command, giving one time base for command and motion. Host machine, OS, interface (USB, CAN, Ethernet) and SDK version recorded.
  • Trials. 50 step commands on the index finger's proximal joint.
  • Reported. Median, 95th percentile and max. Resolution is one frame period. SDK readback latency (time until readback first changes) reported alongside.

Procedure

  1. Hold the finger at rest mid-range for at least 1 s.
  2. Send a fixed-amplitude step (a 20° target change in v0.1) and fire the LED in the same call.
  3. Wait a random 0.5–2 s interval before the next step, so commands do not phase-lock to the hand's control loop.
  4. For each trial, find the LED-on frame and the first frame where fingertip marker displacement exceeds three times its at-rest noise.
  5. Latency = frame difference × frame period.

Known limitations: the number includes the host OS, driver and bus, not the actuator alone. It is valid for the stated stack. LED switching delay is measured once per session and subtracted.

3 · Per-joint backlash deg
  • Definition. Lost motion on direction reversal: how far the commanded joint position travels after a reversal before the link measurably moves.
  • Unit. degrees.
  • Setup. ArUco markers on the two links either side of the joint; joint angle is the relative angle between them. Where a link is too small for a marker, a light rigid pointer carries it; its mass is recorded. Underactuated or coupled joints are measured per actuator.
  • Trials. Each actuated joint, at 25 / 50 / 75 % of its range, 5 reversals in each direction.
  • Reported. Per joint mean ± sd. The card shows the maximum across joints. Where SDK readback disagrees with the optical angle, both are reported.

Procedure

  1. Move the joint slowly (≤ 5°/s in v0.1) in one direction past the test position so the drivetrain is loaded on one side.
  2. Reverse the command in small increments while recording command, SDK readback and optical angle.
  3. Backlash = commanded travel after reversal until the optical angle changes by more than its noise band.
  4. Repeat in the opposite direction, then at the other two positions.

Known limitations: measured unloaded and quasi-static; backlash under grip load can differ. If the joint encoder sits on the motor side, SDK readback will not show backlash at all, which is why the optical angle is the reference.

4 · Max grip force and force curve N
  • Definition. The highest force the hand sustains for at least 1 s on the gauge at maximum commanded effort, and the curve of measured force against commanded effort.
  • Unit. N.
  • Setup. Force gauge fixed to the mount. Two contact cases: thumb–index tip pinch across the gauge with a flat rigid anvil, and single index fingertip pressing the gauge at a fixed pose. Contact geometry and finger pose recorded. Vendor current and safety limits left at default and recorded.
  • Trials. 3 effort sweeps up and down; 5 repeats at 100 % effort.
  • Reported. Max force mean ± sd per contact case; the force curve; deviation from a linear fit; hysteresis between up and down sweeps. The card shows the pinch value.

Procedure

  1. Zero the gauge with the fingers just off contact.
  2. Step commanded effort from 0 to 100 % in 10 % increments, holding 3 s at each; take the mean of the last 1 s.
  3. Step back down to 0 % the same way.
  4. Repeat the sweep three times, then take five 3 s holds at 100 %.
  5. Rest the hand at least 5 min before measurement 5.

Known limitations: force depends on contact geometry and pose, so values hold only for the stated case. This is a short-hold value, not continuous duty (see measurement 6). Whole-hand power-grasp force needs a split-cylinder load cell and is not in v0.1.

5 · Drift after 500 continuous grasps mm
  • Definition. How far each fingertip's landing position moves after 500 back-to-back grasp cycles, compared with before.
  • Unit. mm.
  • Setup. As measurement 1, plus a rigid test object held in the fixture. Object size and material recorded.
  • Trials. One 500-cycle run; before and after capture uses the measurement-1 targets, 10 cycles each.
  • Reported. Centroid shift per finger immediately after the run and again after a 10 min rest. The card shows the worst finger immediately after. Also whether SDK readback reports the same shift.

Procedure

  1. Capture baseline fingertip positions at the three targets.
  2. Run 500 cycles without pause: open, close on the object at 50 % commanded effort, hold 1 s, open. Log cycle time, readback and any SDK-reported temperatures.
  3. Re-capture the three targets immediately.
  4. Rest 10 min powered, then capture again to separate thermal drift from lasting shift.

Known limitations: 500 cycles stands in for a working session, not for lifetime or endurance. Cycle rate depends on the hand's speed and is reported rather than fixed.

6 · Thermal derating after 10 min full load % of cold
  • Definition. Max pinch force measured right after 10 minutes of full-effort gripping, as a percentage of the cold max pinch force from measurement 4.
  • Unit. % of cold value.
  • Setup. As measurement 4, pinch case. Ambient temperature logged throughout. No forced airflow unless the vendor specifies it, in which case it is recorded.
  • Trials. One 10 min load run, then 5 max-force holds within 60 s.
  • Reported. Hot / cold ratio (%), the force trace over the 10 min, and the time and nature of any protective event (current fold-back, fault, shutdown).

Procedure

  1. Start from the rested state after measurement 5.
  2. Grip the gauge at 100 % commanded effort, 2 s on / 1 s off, for 10 min, recording force continuously.
  3. Within 60 s of the end, take five 3 s holds at 100 % effort.
  4. Ratio = mean hot max force ÷ mean cold max force.

Known limitations: the result depends on ambient temperature and airflow, both recorded. Vendor thermal protection is never disabled; if it triggers, that is part of the result and is reported as such.

Environment and conditions

Everything that could move a number is either held fixed or written down.

Conditionv0.1 rule
AmbientTemperature and humidity logged throughout the session. Target 20–26 °C; values outside that are flagged on the card.
Warm-up15 min powered, then 20 full open/close cycles before measurement 1. "Cold" means after warm-up, before any sustained load.
Firmware and softwareFirmware version, SDK version, control mode and gains recorded. Vendor defaults; any vendor calibration routine run as documented and recorded.
MountingRigid fixture at the hand's wrist flange, not on an arm. Hand orientation relative to gravity stated.
PowerVendor-specified supply at rated voltage. Supply voltage and current logged.
InstrumentsForce gauge calibration date and certificate recorded, and checked against reference masses before each session. Camera intrinsics calibrated each session; frame rate verified against a reference clock. Marker sizes measured.
OrderMeasurements 1 → 6 in sequence, with the rests stated above.

Every reported value carries its measurement resolution. Raw logs and video are kept for each session and shared with the vendor.

How this maps to DexBench

DexBench describes dexterity through five Dexterity Regimes: Grasp Diversity, Spatial Precision, Temporal Precision, Contact Precision and Context Awareness. DexBench treats these as abilities of a whole system, not of a hand. A card measures the hand's physical limits, which bound what any system built on it can reach. The table shows which regime each measurement helps explain. It is not a DexBench score.

Card measurementDexBench regime
Fingertip repeatabilitySpatial Precision
Command-to-motion latencyTemporal Precision
Per-joint backlashSpatial Precision; Contact Precision
Max grip force and force curveContact Precision; partly Grasp Diversity (which objects can be held)
Drift after 500 graspsSpatial Precision, over a session
Thermal deratingContact Precision, over time
—Grasp Diversity is only partly covered: v0.1 does not measure kinematic range or contact surfaces.
—Context Awareness has no hardware measurement. It is a property of perception and control, not of the hand.

DexBench is an independent specification published by RLWRLD, with NVIDIA listed as a platform partner. Robotics Center is not affiliated with DexBench, RLWRLD or NVIDIA; the mapping is our own. Regime names as published at dexbench.org, retrieved 2026-09-27.

Card format and versions

Each card is one measurement session on one unit.

Comment on the draft

We want this reviewed by the people who build and use these hands. Tell us what is wrong, missing or unfair to a particular design. Email contact@roboticscenter.ai with the subject "[Dexterity] Protocol v0.1 comment".

Comment on the draft About the Dexterity Card

Changelog