Industry

Before asking whether a cobot station is tiring, ask where operators hesitate: one depth camera splits workload into five sources, and added collaboration raises mental and time pressure first

An assembly line has just added a collaborative robot. The robot hands over parts and tools; the operator assembles. A month later output is barely up, and the team lead reports that operators feel more tired. Ask where, and nobody can say. The work is not heavier. Their heads are busier, and their hands keep waiting.

A smart manufacturing team at HKUST (Guangzhou) (Xiong, Guo, et al., September 2026) offers a checkable way to split that feeling. Operators wear no sensors. One RGB-D depth camera plus the robot’s own state data breaks workload into five visible sources and shows them live on a screen. As collaboration steps increased, the first things to rise were mental demand (7.33 to 10.14) and temporal demand (5.55 to 7.30). Physical demand did not change significantly. This note maps plant management moves. It is not a reprint.

What the paper reports

The problem: in human-robot assembly, operators must do manual work while tracking robot motion and task information. NASA-TLX questionnaires come only after the task and miss changes during it. ECG and skin-conductance signals need wearables, are awkward on a shop floor, and are noisy across individuals.

Method and results (as stated in the paper):

• Station zones. The shared workbench is calibrated into assembly, storage, instruction, and robot-collaboration areas. The camera tracks head orientation, upper body, and hands, and a temporal confirmation step filters brief flickers before deciding where the eyes are and what the hands are doing.
• Five indicators. Assembly execution demand (AED), tool and part identification demand (TID), instruction checking demand (ICD), context-aware attention switching demand (ASDctx, weighted by where each switch starts and ends rather than a raw count), and overall hesitation demand (OHD, the share of robot-collaboration time with sustained low hand motion). They combine into one composite score, but each stays visible so a change can be traced to its source.
• Study design. A UR5e cobot replayed taught trajectories to deliver parts while participants assembled a simplified gearbox (base, cover, shafts, gears, bearings, fasteners). Level 1: the robot delivered one part. Level 2: two parts plus a screwdriver. Level 3: two parts plus a fastening module of screwdriver and screws. Ten participants (mean age 21.6±2.0), ethics-approved.
• Perceived workload. Raw NASA-TLX rose 6.75 → 7.26 → 8.45 (Friedman χ²(2)=6.20, p=0.045). Mental demand rose 7.33 → 8.31 → 10.14 (p=0.035). Temporal demand rose 5.55 → 6.51 → 7.30 (p=0.033). Effort (7.40 to 9.74) and frustration (4.37 to 7.10) rose but not significantly. Physical demand (p=0.247) and self-rated performance (p=0.607) showed no significant change.
• Physiological check. ECG, EMG, and EDA served only as independent references. In 7 of 9 participants with complete data, the vision-derived composite was significantly associated with ECG features. Detected hesitation episodes aligned in time with skin-conductance responses and muscle activity.
• Live deployment. The pipeline ran in real time at the station with no sensors on the operator. The interface shows all five indicators beside the composite score.
• Stated limits. Small sample, one calibrated setting, fixed weights. The output is a behavioral proxy, not a direct physiological measure. Adapting the robot from the indicators (slower motion, different handover timing, a new task split) is proposed but not yet tested.

Source: arXiv:2609.15232 (PDF). Title: A Vision Based Framework Integrating Attention and Action Cues for Interpretable Cognitive Workload Assessment in Human-Robot Collaborative Assembly.

For the plant: ask where the load sits, then decide what to change

Many plants judge a new cobot station on two things: did takt improve, and do operators say they are tired. This study’s numbers point somewhere more specific. As collaboration steps grow, physical load barely moves, while mental and time pressure climb. Operators are watching the robot, waiting for handovers, hunting for tools, and checking instructions. That load is invisible, yet it slows takt and adds errors. A minimal plan:

1. Pick one cobot station to diagnose. Choose one where output is below plan or operators say it feels harder without saying why.
2. Record where eyes and hands are, by zone. You do not need a camera system on day one. A team lead with a stopwatch and a simple sheet can log, over one shift, time spent searching for parts, checking instructions, watching the robot, and waiting with hands still. Find the largest share first.
3. Match one fix to one source. For part searching, change part presentation and color-code bins. For instruction checking, rewrite instructions as one step per picture, placed straight ahead. For robot watching, fix the handover position and rhythm. For hands pausing at handover, slow and smooth the robot’s delivery, or hand that step back to the operator.
4. Re-measure the same indicator after the change. Check whether that share dropped, not only total output. Use the data to redesign stations and takt, not to score individuals, and operators will help.

The organizational work here is turning “the operators say it is tiring” into a station problem you can locate, fix, and re-check, so the cobot investment shows up as real labor productivity.

A fifteen-minute self-check

At your newest cobot station, has anyone measured how long, per shift, operators stand with their hands still waiting for the robot? If operators say the job got harder, can you tell whether it is part searching, instruction checking, or robot watching? If not, the next thing to fix is not the robot. It is the station diagnosis.

Figures and wording follow the arXiv paper; this note only maps plant management moves.