When Hospital Robots Ask Staff for Help: Forecasting Human Support in Robot Dispatch
# Raw simulation outputs for "When Hospital Robots Ask Staff for Help: Forecasting Human Support in Robot Dispatch" **Krzysztof Siwek** (Warsaw University of Technology) and **Aleksandra Świetlicka** (Poznań University of Technology) This record contains the raw per-run outputs, frozen experiment designs, calibration logs, and summary analyses of the discrete event simulation study reported in the article above (submitted to the *International Journal of Social Robotics*). The simulation code is published separately on GitHub (repository `hospital-robot-staff-support`); together, code and data reproduce every table and figure of the article. ## Study in brief A discrete event simulation replays 33,079 recorded transport requests of the University Medical Center Freiburg (Hess and Stampa, 2026, https://doi.org/10.5281/zenodo.19085901) and compares human transport with robot dispatch rules when robot missions need staff for loading, receiving, and cleaning. The main experiment crosses three support arrangements (mobile transporters, stationary local workers, mixed), three levels of staff unavailability (b = 0, 0.4, 0.8), and two reporting delays (0 and 10 min), giving 18 conditions, with 20 paired seeds (103000 to 103019), seven warm-up days, seven measured days, and one clearance day per run. Later phases add random admission controls at matched robot coverage, a staff time front, an optimizer volume control, a threshold rule that compares the forecasts at equal coverage, exploratory operating scenarios, a class mix with 20% instead of 40% urgent specimens, and an availability gate that admits a robot only when idle workers can serve it. The run files also hold the staff-side measures of the article (support calls, call-to-arrival time, robot waiting time, displaced other duties, overtime). Task classes, deadlines, service times, and staffing are model assumptions. ## Files | Archive | Content | Run files ||---|---|---|| `00_designs_calibration_and_summaries.tar.gz` | Frozen designs (`design_*.json`), calibration and verification logs, model inputs, `main_runs_subset.csv`, summary analyses (`analysis_*`) | – || `01_main_experiment.tar.gz` | Main experiment: human transport, staff rollout, omitted support, own completion, optimizer | 1,800 || `02_controls.tar.gz` | Matched random, idle pair and idle surplus rules; calibration runs | 1,080 + 288 || `03_front_allowance_rules.tar.gz` | Staff time allowance rules (φ = 3, 10; seeds 103000–103007), pilot; sensitivity (4 and 8 robots, robot speed, support time) | 576 + 288 + 816 || `04_optimizer_volume.tar.gz` | Random arms at the coverage of the optimizer and of staff rollout; calibration and verification | 552 + 1,072 || `05_omitted_volume_and_mission_logs.tar.gz` | Random arm at the coverage of omitted support; logged reruns of staff rollout, omitted support and matched random; calibration and verification | 360 + 1,080 + 432 || `06_front_volume_and_threshold_rule.tar.gz` | Allowance rules on seeds 103008–103019, four random arms along the front, threshold rule at the coverage of staff rollout (`results/frontvol`); calibration, verification and test runs | 2,664 + 3,890 || `07_scarcity_unattended.tar.gz` | 53 and 44 transporters, unattended receiving; random arms | 1,536 || `08_equal_budget.tar.gz` | Three or six local workers within the same staffed hours; random arms, roster pilot | 1,064 || `09_operating_levers.tar.gz` | Demand ×1.1 and ×1.2, night-only operation, multi-load tours, combined scenario; random arms, probes | 4,268 || `10_warmup_check.tar.gz` | Fourteen instead of seven warm-up days (three conditions, two policies, 20 seeds) and the zero-probability random arm | 121 || `11_class_mix.tar.gz` | Class mix with 20% specimens (human transport, staff rollout, omitted support, matched random, threshold rule; 18 conditions, seeds 103000–103009, `results/specmix/main`); greedy, threshold-grid and calibration runs on pilot seeds (`results/specmix/calib`); `design_specmix.json`, `specmix_calibration.json`, `analysis_specmix/` | 900 + 1,512 || `12_availability_gate.tar.gz` | Availability gate on staff rollout (18 conditions, seeds 103000–103019, `results/gate`) and two reproduction tests with the gate disabled; `design_gate.json`, `analysis_human/` (staff-side measures of all phases) | 360 + 2 || `SHA256SUMS.txt` | Checksums of all archives | – | Pilot, calibration, verification, probe, and smoke runs use seeds from 101000 to 104907 and were never used for inference. Version 1.1.0 adds archives 11 and 12 (2,774 run files) to the 21,887 run files of version 1.0.0; the archives 00 to 10 are unchanged. ## Run files Each run is one JSON file named `<tag>__<condition>__<policy>__<seed>.json` (the main experiment omits the tag), where `<condition>` is `<arrangement>_b<unavailability in %>_lag<reporting delay in min>`, for example `mobile_b80_lag0`. Each file contains the configuration of the run (`workflow`, `warmup_days`, `measurement_days`, `seed`) and its outcomes. The main fields are: - `weighted_tardiness_per_request`: priority-weighted tardiness per measured request, $J$ (weighted minutes);- `late_fraction`: share of measured requests completed after their deadline;- `robot_coverage`: share of requests completed by robots;- `net_released_minutes`: manual transport time avoided minus staff time for robot support, over the seven measured days;- `prediction_mae`: mean absolute error of the dispatcher's completion forecast;- `created`, `completed`, `incomplete`, `robot_completed`, `human_completed`: request counts;- `mission_log` (logged phases only): every measured robot mission with class, deadline, deterministic route, elevator rides, shift, queue and idle workers at the decision, and weighted tardiness, plus the split of tardiness into robot-served and human-served requests;- `human_support_tasks`, `human_support_wait_minutes`, `*_support_queue_wait`, `*_support_approach`, `*_displaced_background_minutes`, `*_overtime_minutes`: staff-side measures (calls, call-to-arrival time and its parts, other duties displaced by transport and support work, overtime), with `*` = `mobile` or `local`;- `gate_blocked` (availability gate only): measured decisions in which the gate rejected the robot chosen by staff rollout;- `quota_margins` (threshold rule only): share of tardiness ties among the decisions in which the forecast without support prefers a robot;- optimizer diagnostics (`optimizer_*`) for runs with the CP-SAT scheduler. Seeds share the demand and unavailability streams across policies, so differences between policies on the same seed and condition are paired. Runs of the heuristic policies are deterministic given the seed; optimizer runs depend on wall-clock solver limits. ## Reproduction Unpack the archives into the root of the code repository (they create `results/<phase>/` and the design and analysis files there), then run the analysis scripts in `ctl/`, for example `python ctl/analyze_frontvol.py`, `python ctl/analyze_specmix.py`, or `python ctl/analyze_gate.py`. The runners skip existing runs and refuse to mix runs from a changed design. ```bashsha256sum -c SHA256SUMS.txtfor f in *.tar.gz; do tar xzf "$f" -C /path/to/hospital-robot-staff-support; done``` ## License The simulation outputs in this record are released under the Creative Commons Attribution 4.0 International license (CC BY 4.0). The underlying request and pathway data belong to the Freiburg dataset and are subject to its own license. ## Citation Please cite the article (details will be added on publication), this record, and the Freiburg dataset.
Authors
- Krzysztof Siwek (ORCID: https://orcid.org/0000-0003-2642-2319)
- Aleksandra Świetlicka (ORCID: https://orcid.org/0000-0002-2576-7625)
Institutions
- Warsaw University of Technology (PL)
- Poznań University of Technology (PL)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.5281/zenodo.24310954
- Primary Topic
- Healthcare Operations and Scheduling Optimization
- Type
- article
- Field-Weighted Citation Impact
- 0.00