Answers are grounded in this paper and cite their sources.
Although contactless hand-gesture interfaces can support intuitive human–robot interaction, transient classification errors may trigger unintended commands when classifier outputs are mapped directly to robot actions. This study presents a processing framework for contactless control of an ABB IRB 120 industrial robot. The framework combines RGB acquisition, geometric normalization of 21 MediaPipe hand landmarks, population-based sample filtering, multilayer perceptron classification, temporal transition stabilization, and a two-hand command grammar. Thirteen gesture classes were represented by 63-dimensional landmark vectors, and 19,250 samples were retained after processing. The classifier achieved an approximate mean per-class recall of 99.46% on the held-out test subset. Four temporal processing configurations were compared using 30 repetitions per interface action: raw output, majority voting, three-consecutive-label confirmation, and the proposed cellular-automaton (CA)-based stabilizer. The latter two methods improved command reliability relative to the raw output and produced comparable results. The CA stabilizer processed labels sampled every 100 ms and introduced a validation delay of approximately 300 ms. Robot-program control achieved an 86.7% success rate across 90 trials, while Cartesian direction control achieved a 97.2% success rate across 180 trials. The mean displacement overshoot was 2.42 mm, or 24.2% of the nominal 10 mm command. The framework provides a practical perception-to-control interface while complementing, rather than replacing, certified robot safety functions.
No takes yet. Share an insight, caveat, or question.
Bieńkowski et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: