Winning Huawei’s Global Smart Port Challenge: Key Strategies and AI Techniques

Competition Overview

In 2024 the 10th Huawei Software Elite Challenge – the "Planck Project" – held a global final that attracted more than 8,000 teams from over 800 universities. The Harbin Institute of Technology team "Yuan Dream Star" won the championship.

Motivation

The author joined the contest to challenge himself after seeing the previous year’s SimpleDemo and champion team, hoping to beat the champions and push his own limits.

Problem Description

The task simulates a smart port, a key development direction for city trade ports. Using Huawei Cloud’s AI platform, participants must provide algorithms for intelligent cargo handling, ship docking, and cost reduction to maximize efficiency and lower unloading costs.

Solution Overview

The solution is organized into three competition phases – preliminary round, semi‑final, and final – each requiring specific code modifications and algorithmic enhancements.

Preliminary Round

We reused the SimpleDemo framework, customizing the robot decision function and ship avoidance algorithm. The robot decision assigns a sell task to every fully loaded robot and a buy task to all robots based on a value‑to‑time ratio, while also considering item disappearance time, target stability, and negative profit for unsellable items.

Ship decision also splits into sell (choose the nearest delivery point) and buy (assign a berth with the most cargo). The ship never changes its target mid‑flight.

These modifications earned us first place in the national preliminary round.

Semi‑Final Round

The semi‑final introduced robot and ship purchasing mechanisms and required autonomous ship navigation and avoidance. We enhanced ship decision by adding purchase logic, dynamic buying factors, and refined avoidance strategies. Robot decision remained focused on optimal buy/sell assignments.

Ship avoidance now attempts to find a collision‑free path of equal minimal distance; if none is found, the ship continues on its current path until a collision threshold triggers local avoidance, selecting the best of 25 candidate points.

Final Round

The final expanded the map, added a large‑model Q&A module, and introduced multi‑team confrontation. Performance became critical due to the massive number of cargo items.

We optimized robot decision by pruning low‑value items (processing only the top 25% when decision time exceeds 4 ms) and added special handling for the new Q&A state. The item disappearance value was removed, and the same‑target factor was reduced to 0.5.

Ship navigation switched from simple Dijkstra to a BFS optimized with an ordered dictionary (depth → stack) to handle the larger map efficiently.

We also implemented advanced avoidance for opponent robots and ships, assuming moving opponents will yield to us, and locking positions for a few frames to search collision‑free paths.

The large‑model module runs in a background thread, constantly polling for questions, using greedy decoding (top_k=1) and a short English suffix prompt to minimize token output, achieving competitive response speed.

Overall Workflow and Definitions

Each simulation frame performs robot actions (decision, pathfinding, avoidance), ship actions (decision, pathfinding, avoidance), and purchase decisions for both. In the preliminary round only robot and ship decisions are needed; the semi‑final and final also consider purchases.

Key definitions include Workbench (item location), robot buy/sell decisions, ship buy/sell decisions, and combined purchase strategies (static minimum numbers plus dynamic factor‑based buying).

Conclusion

The competition presented many challenges, but through systematic algorithm design, data‑structure selection, and performance tuning, the team secured the championship while gaining deep insights into AI‑driven logistics optimization.