Answers are grounded in this paper and cite their sources.
The construction industry has a higher accident rate than other industries and a shrinking inflow of new workers, and obstacle avoidance is a core task for autonomous construction equipment. Many earlier studies treated every obstacle as a single category with an identical penalty, which limits their use on real sites. This study proposes a deep reinforcement learning (DRL) reward function that reflects the quantitative criteria of construction safety regulations. RecurrentPPO, combining Proximal Policy Optimization (PPO) with Long Short-Term Memory (LSTM), served as the local planner and Hybrid A* as the global planner. From KOSHA GUIDE C-48-2022 and the Rules on Occupational Safety and Health Standards, type-specific obstacle penalties, a four-zone safety-distance scheme and speed terms were designed. Three groups and three reward ablations were compared in simulation. Against a baseline reward with a single safety-distance term, the difference in success rate (71.2% vs. 58.9%) was not statistically significant, and the drop in collision rate from 41.0% to 25.1% was significant in Welch’s t-test but only marginal in the Mann–Whitney U test. The proposed reward had its clearest effect near workers. The share of steps above the 10 km/h limit fell from 90.9% to 51.5%, and worker collisions were halved. A safety-regulation-based reward thus mainly shapes regulation-consistent driving rather than raising the arrival rate.
No takes yet. Share an insight, caveat, or question.
Kim et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: