minphak/Demo_C2-APP-013

AI20K-170 — Synthetic-to-Real Dataset for Tabletop Robot Object Detection

Sim-to-real YOLO dataset built for tabletop robot perception on 10 YCB objects. Generated with PyBullet simulation, then refined with generative realism (SD1.5 + ControlNet or Flux.2-Klein) and validated against real photos held out from training.

Project repo: https://github.com/ (Team 013, AI20K-170, Cohort 2)

Contents

• datasets/raw/ — real YCB reference photos used as the lab_raw baseline
• sim/, golden/ — PyBullet-rendered frames (RGB, depth, canny, YOLO labels)
• golden_2_report_for_lab_sim_sd15/ — Golden-2 benchmark data: 10-class multi-object comparison (lab_raw / sim / realism SD1.5)
• golden_3/ — Golden-3 benchmark data: single-class power_drill ablation (lab_raw / sim / realism Flux.2-Klein), training runs, weights, and the 30-photo real-life eval set
• YOLO26_SimToReal_Benchmark_Report.md — Golden-1 report (7-class, historical)
• YOLO26_Golden2_Benchmark_Report.md — Golden-2 report (10-class, multi-object)
• YOLO26_Golden3_PowerDrill_Benchmark_Report.md — Golden-3 report (power_drill ablation)

Benchmark summary

Run	Scope	Variant	mAP50	vs sim
Golden-2	10 classes, multi-object	lab_raw	0.762	ref
Golden-2	10 classes, multi-object	sim	0.912	baseline
Golden-2	10 classes, multi-object	realism (SD1.5)	0.949	+3.7%
Golden-3	power_drill only	lab_raw	0.018	ref
Golden-3	power_drill only	sim	0.023	baseline
Golden-3	power_drill only	realism (Flux.2-Klein)	0.802	+77.9pt

All variants share identical YOLO26n hyperparameters, seed 0, 50 epochs; only the training data differs. Evaluation is always on real photos held out from training.

Object classes (YCB)

cracker_box, mustard_bottle, tomato_soup_can, mug, bowl, apple, master_chef_can, sugar_box, power_drill, bleach_cleanser.

License

MIT.