How Our Team Won the Huawei Software Elite Challenge with Advanced AI Path‑Planning Strategies

Background and Motivation

In the 2023 9th Huawei Software Elite Challenge "Planck Project" global finals, nine teams shared a total prize pool of 660,000 CNY. The Central South University team, named "Why Is the Competition Period So Long", won the runner‑up position. Team member Tang Jing wrote this report to share the team's motivation, problem‑solving approach, and outcomes.

Problem Statement

The final task was more complex than previous editions, demanding not only high‑level algorithmic skills but also intricate detail implementation, including scheduling, movement, defense, and attack strategies.

Solution Overview

The solution consists of four main components: scheduling, work‑point splitting, map point sampling, and robot navigation.

Scheduling (Version 8)

The team defined a production line (line 7) composed of a series of workstations that can fully produce and sell item 7. Based on analysis of previous versions, the scheduling strategy aims to:

Balance production of items 4, 5, 6 on line 7.

Follow the "carry‑along" principle: when selling an item, bring a material for the next item.

Avoid congestion by preventing too many robots from heading to the same workstation.

Prefer not to sell line‑7 items to workstation 9.

Prioritize filling material slots of workstations that already have some resources.

Maximize profit per unit of robot movement.

The core scheduling code scores each potential task by profit / (distance + 10) and adjusts weights according to the rules above, allowing robots to select the highest‑scoring task.

Work‑Point Splitting

To improve robot precision at workstations, the team sampled points within a 0.4‑meter radius around each workstation at eight directions, spaced 0.05 m apart, and grouped them into five sets for fine‑grained navigation.

Map Point Sampling

The robot radius leads to a 201 × 201 grid where each intersection is a sampling point. To reduce computational load, the team performed stride sampling and special‑case handling, keeping at most ten thousand points (typically a few thousand).

Robot Navigation

Navigation combines three layers:

Global path planning: shortest‑path preprocessing for the initial stage.

Local navigation: A* search each frame to compute a path to the target, with distance penalties for enemy robots and obstacles.

Collision avoidance: DWA was used in the preliminaries, but ORCA proved more effective in the finals, especially against moving enemy robots.

When A* fails to find a path, a fallback local navigation module provides a safe intermediate goal.

Attack and Defense Strategies

Two types of offensive robots chase enemies to potentially block enemy transport robots. Defensive robots occupy enemy workstations on line 7 to disrupt their production line. The team also introduced a "rushAble" flag to classify workstations (corner, risky, or safe) and dynamically disable tasks that could lead to deadlock.

Key Takeaways

The team faced many setbacks, including a massive bug count just before the finals and last‑minute strategy revisions. Nevertheless, intensive overnight coding sessions, rapid prototyping of ORCA, and iterative tuning enabled them to advance to the finals and secure the runner‑up prize.

— Tang Jing, Central South University "Why Is the Competition Period So Long" team