A dexterous hand sample has arrived. The next seven days decide whether it earns a place on your platform or goes back in the box. Most teams evaluate by playing with it for a few days and going with gut feel — which is exactly how costly late-stage surprises happen.
The point of a sample evaluation is not to prove the hand works. It is to find out how it fails, and whether those failure modes are acceptable for your application. Here is the protocol we use when qualifying hands, with a gate at the end of every day. Fail a gate and you stop.
Day 1 — Unboxing and physical inspection
Before you connect power or a bus, inspect the hardware. Mechanical problems found here save hours of debugging later.
| Check | Pass | Fail = stop |
|---|---|---|
| Weight vs datasheet | Within ±5% | >10% deviation suggests the wrong configuration was shipped |
| Joint free movement (power off) | Every joint moves smoothly by hand | Any joint that binds or grinds is a mechanical defect |
| Connectors | Clean, no bent pins, locks positively | Bent pins or a loose fit means an unreliable link |
| Documentation | Quick-start guide with a clear wiring diagram | No documentation means painful integration |
| Accessories | Power supply, cables, adapter, mounting bracket present | Missing parts — contact the supplier before powering on |
Photograph all sides before you start. That is your baseline if you ever need to claim warranty. And do not power on a visibly damaged hand.
Day 2 — Communication and basic control
Goal: bidirectional comms and one joint moving. If you cannot talk to the hand by end of day two, integration will be a long fight.
- Identify the protocol — CAN, CAN FD, EtherCAT, RS485 or serial — from the datasheet and pinout.
- Set up the bus interface: USB-to-CAN adapter, EtherCAT master, or serial converter.
- Install drivers and SDK. Flag it if the only instructions are in Chinese — that is a support risk worth knowing about before you order production units.
- Power on at the specified voltage (commonly 24 V) and watch the indicators.
- Send a heartbeat, read all joint positions, then command one joint to move 10°.
- Test emergency stop — you need to be able to halt motion instantly.
| Symptom | Likely cause | First thing to try |
|---|---|---|
| No response at all | Wrong baud rate or node ID | Check datasheet defaults; try 1000 kbps then 500 kbps |
| Some joints respond, some don't | Address conflict or broken daisy chain | Check wiring continuity, verify node IDs |
| Jittery motion | Gains untuned or bus noise | Check grounding, lower PD gains |
| SDK examples crash | Library or OS mismatch | Ask the supplier for a known-good version |
| Warm at idle | Firmware bug or wiring fault | Stop immediately — hardware risk |
Day 3 — Motion characterisation and calibration
Now measure how it actually moves. Everything downstream depends on these numbers.
| Parameter | How to measure | Good result |
|---|---|---|
| Range of motion | Command min to max, measure the real angle | Matches datasheet ROM |
| Position accuracy | Five target angles per joint, commanded vs actual | ±0.5° |
| Repeatability | Same angle ten times, measure spread | ±0.1° on precision hands |
| Maximum speed | Time a full flex–extend cycle | Meets datasheet speed |
| Step response | Step input, time to 90% of target | Under 100 ms |
| Backlash | Approach one position from both directions | <0.5° direct drive, <2° tendon |
Accuracy more than 2° off spec is either a calibration problem (fixable) or an encoder problem (not fixable). Ask the supplier which before you proceed. If your application needs grasp planning, also map joint angles to fingertip position: command the hand to touch known reference points on a grid, record the angles, fit a kinematic model, then verify at five points you did not use for fitting.
Day 4 — Grasp testing and force measurement
This is where the spec sheet meets reality. Build one standard object set and use it for every hand you evaluate, so scores are comparable.
| Grasp type | Test objects | What it tells you |
|---|---|---|
| Power (cylindrical) | Bottle, can, 50–80 mm cylinder | Maximum grip force and wrap-around |
| Precision (pinch) | Pen, 5–15 mm cylinder | Fingertip dexterity and fine force control |
| Lateral pinch | Card, thin book, phone | Thumb-to-finger lateral force |
| Tripod | Pen in writing position, small ball | Three-finger coordination |
| Hook | Bag handle, rope | Load-bearing capacity, finger strength |
| Delicate | Egg, ripe fruit, paper cup | Force regulation and tactile feedback |
With a force gauge or even a kitchen scale: measure maximum fingertip force, force resolution across five commanded levels, force repeatability over ten trials, and — if the hand has tactile sensing — whether slip is actually detected when you load a grasped object until it moves.
Day 5 — Durability and thermal
A hand that works for five minutes and drifts after thirty is useless. Program a repetitive open–grasp–release cycle and run it at least two hours, logging joint temperature every fifteen minutes.
| Metric | Good | Acceptable | Failing |
|---|---|---|---|
| Motor temperature after 1 h | <45 °C | 45–60 °C | >60 °C — thermal shutdown risk |
| Grasp success rate, start vs 1 h | No change | <10% drop | >10% drop |
| Position drift | <0.5° | 0.5–1° | >1° — thermal growth or encoder drift |
| Audible noise | No change | Slight change | New grinding or clicking = wear |
Then run 1,000 grasp cycles. Stable at 1,000 usually means 10,000–50,000 before meaningful wear. Degradation before 1,000 means you should plan for frequent maintenance.
Day 6 — Integration with your platform
The hand stops being a bench device today, and this is the day that exposes integration problems.
- Mount it — check the wrist flange pattern, that the weight sits inside arm payload, and that nothing collides at the extremes of wrist rotation.
- Route the cable through the arm's internal channel and check for pinching at joint extremes.
- Power it from the robot bus and verify voltage stability under load.
- Write the driver into your control framework (ROS 2 or otherwise) and calibrate the tool centre point.
- Run an end-to-end test: move the arm to a target, grasp, move away.
The most common failure here is cable management: a cable that behaves perfectly on the bench pinches or fatigues once the arm sweeps its full range. Test every configuration, not just the home pose. Also verify your safety watchdog opens the hand on comms loss rather than freezing it closed, and that one e-stop halts arm and hand together.
Day 7 — Score it and decide
| Category | Weight | Score 1–5 |
|---|---|---|
| Mechanical quality | 20% | |
| Motion performance | 20% | |
| Communication reliability | 15% | |
| Grasp capability | 15% | |
| Durability | 10% | |
| Documentation quality | 10% | |
| Integration effort | 10% |
Weighted total then maps to a decision: 4.0–5.0 go — order production units and spares. 3.0–3.9 conditional — acceptable with named caveats to fix before scaling. 2.0–2.9 hold — request manufacturer fixes or evaluate alternatives. Below 2.0 no-go — return it.
Run the protocol on a real sample
Tell us the grasp types you care about, the payload and the protocol, and we will shortlist hands across our partner factories and quote the sample and the production price together. Start from the Linker Hand range; if teleoperation or demonstration data is part of your plan, evaluate a motion capture glove in the same window — same SDK, same ROS packages.
Specs quoted from manufacturer datasheets; final figures confirmed at quotation.