Open Access

ARTICLE

SW-DWNS: A Single-Wave Autonomous Navigation System in Partially Observable, Highly Dynamic Warehouses

Xianhui Fan¹, Zongwei Li^1,*, Yuxuan Zhai¹, Zhenyu Li²

1 School of Economics and Management, Shanghai Institution of Technology, Shanghai, China
2 School of Cultural Heritage and Information Management, Shanghai University, Shanghai, China

* Corresponding Author: Zongwei Li. Email:

Computers, Materials & Continua 2026, 88(3), 76 https://doi.org/10.32604/cmc.2026.083491

Received 04 April 2026; Accepted 11 June 2026; Issue published 23 July 2026

Abstract

Single-wave order picking in dynamic warehouses is a sequential multi-goal navigation problem. A robot must visit an ordered set of shelves and then a delivery station while avoiding moving obstacles under partial observability. Existing approaches either entangle long-horizon task logic with low-level obstacle avoidance or rely on static-environment assumptions that limit responsiveness in dynamic settings. This paper proposes the Single-Wave Dynamic Warehouse Navigation System (SW-DWNS), a lightweight scheduling framework that extends a pretrained ColorDynamic point-to-point local planner to ordered warehouse picking without retraining. The scheduler maintains a shelf queue, exposes only the active subgoal to the local policy, and advances the queue through an arrival-gated state machine with a geometrically derived pickup threshold. We evaluate SW-DWNS in a controlled dynamic warehouse simulation benchmark with three layouts, pickup horizons K ∈ {1, …, 5}, stochastic obstacle motion, sensor noise, and obstacle-size randomization. Across the full layout–horizon grid, SW-DWNS achieves 87.3% average success, compared with 26.8% for A* search combined with the Dynamic Window Approach (A* + DWA) and 1.3% for ColorDynamic-only. The gain is not obtained by sacrificing safety: the average collision rate is reduced from 69.0% for A* + DWA to 7.5% for SW-DWNS. Factor-wise ablation shows uneven robustness across disturbances: sensing noise and obstacle-size randomization have limited impact (main effects of +1 and +4.5 percentage points), whereas disabling obstacle motion costs 33 percentage points and shifts failures from collision to timeout. A pickup-threshold sweep further shows that reported success rates become inflated outside the admissible geometric band ε ∈ [14, 31] cm. These results show that a decoupled scheduling layer can make a pretrained dynamic local planner effective for sequential warehouse picking, while also revealing the conditions under which that extension remains valid.

---

Keywords

---

Supplementary Material

Supplementary Material File

---