Gilbreth: AI motion study from a workstation video

Manual work is still improved with a century-old tool, the bimanual chart: what each hand does, second by second, coded as operation, transport, hold or delay (ILO symbols). It shows a hand that only holds while the other works. Charting a 4.7-minute video by hand takes about an hour (our estimate). Gilbreth turns a phone video into a draft chart (Excel), method alerts with the fix, and an improvement plan in seconds saved per unit. MediaPipe tracks the hands on CPU; Gemma 4 31B on AMD GPUs (AMD Token Factory) writes its own operation sheet and codes every second from marked frames and hand crops. The engineer corrects the first seconds; a small model learns from them, and a check (the gate) lets it recode the rest only if it beats Gemma on reviewed seconds it was not fitted on. Accuracy = share of hand-seconds (one hand during one second) coded like the human labels; always answering the single best code scores 42–47 %. Each test had its rule written before looking and was scored once, on recordings unused in development. KIT Bimanual (research videos of workshop and kitchen tasks), two new people: Gemma alone 55.4 → Gilbreth 65.1 % and 51.5 → 57.4 %. IMPACT (a second dataset, gearbox work), 8 new runs: 66.5 → 71.1 %. Where coworkers are already labelled, station mode also learns from them: 83.4 % on that second KIT person; at IMPACT it added nothing. Gemma on AMD cost about US$0.008 per minute of video. For a 284-second LEGO assembly the plan estimates 215 s per unit (−24 %), mainly from a fixture that frees the left hand, which holds or waits 222 s while the right hand works. An estimate to confirm at the station. If Gemma is not served, the demo falls back to the best of six models tested on AMD and says so. In the app you can chart a 20-second sample or your clip on AMD, correct its first seconds and see the gate decide how the rest is coded. Code: MIT; the IMPACT-derived hand-motion data the small model starts from is a separate, non-commercial download.