Zero-shot Adaptable Task Planning for Autonomous Construction Robots

🚧 Note: This project is currently under review in a peer-reviewed academic journal. Stay tuned for the official publication and full release of all project materials and detailed results.

Project Overview

This project investigates zero-shot adaptable task planning for autonomous construction robots by employing lightweight, open-source foundation models (LLMs and VLMs) in single- and multi-agent AI architectures. It aims to address challenges such as limited adaptability, high costs, and rigid task-specific programming prevalent in current construction robots.

Key Features

• Zero-shot Task Planning: No task-specific pre-training or detailed human commands required.
• Multi-agent Collaboration: Single-agent and multi-agent (2, 3, and 4 agents) designs to enhance adaptability.
• Cost-Effectiveness: Leveraging lightweight local models instead of expensive cloud-based models.
• Quadruped Robot Integration: Unitree Go2 Edu robot with robotic arm and sensors for real-world experiments.
• Roles Tested: Painter, Safety Inspector, Floor Tiler, each representing varying complexity and dynamism in tasks.

Experiments Conducted

Three critical tasks representing diverse construction activities:

Task	Category	Complexity
Painting	Craft	High variability
Safety Inspection	Non-routine	High variability, low analyzability
Floor Tiling	Craft	Moderate-high variability

Selected Findings

• One of the designs significantly outperformed other architectures and even GPT-4o across several metrics.
• Cost Efficiency: Four-Agent model was 10x cheaper than GPT-4o, making it highly suitable for industry applications.

Contribution & Impact

This research represents a significant advancement toward truly autonomous, adaptable, and cost-effective robotic systems for the construction industry. It sets a benchmark for future research on scalable automation and AI-based task planning across complex, dynamic environments.