TL;DR
This paper introduces a "work companion rover" designed to assist human carpentry workers by autonomously delivering tools and materials, and accompanying them on-site. Leveraging reinforcement learning, the robot navigates complex construction environments and adapts to worker behavior, aiming to improve safety and workflow fluency without replacing human expertise.
What Problem Does This Paper Address?
The construction industry is highly labor-intensive and operates in dynamic, cluttered environments. While robotics has been explored for automating specific construction tasks, traditional approaches often overlook the complex, variable, and skilled nature of human work. Existing robotic solutions struggle with on-site integration due to their lack of adaptability. This paper addresses the challenge of creating assistive robots that can seamlessly integrate into existing construction workflows, specifically focusing on carpentry, to reduce physical strain, mitigate injuries, and enhance efficiency by acting as a supportive "work companion" rather than a fully autonomous replacement.
Methods
The core of the paper's approach is a human-centered design combined with a Reinforcement Learning (RL)-driven modular framework. Key methods include:
- Human-Centered Design: Extensive on-site observations and worker interviews informed the design of the "work companion rover" and its functionalities.
- Modular Framework: A flexible system built on the ROS navigation stack, comprising modules for:
- Hardware System Design: Utilizing a robust Clearpath Husky robot base for navigating uneven terrain, equipped with a multi-modal sensor suite (3D LiDAR, 2D LiDAR, RGB-D cameras, IMU) for perception and localization.
- Mapping and Localization: Employing a recent LiDAR Odometry and Mapping (LOAM) method for 3D point cloud generation, projected into a 2D grid map to balance accuracy and efficiency. An Extended Kalman Filter (EKF) integrates multiple sensor inputs for robust state estimation in dynamic environments.
- Worker Detection and Tracking: Using YOLOv7 for worker detection and depth information for position estimation, with two tilted RGB-D cameras to overcome limited FoV. Multi-Object Tracking (MOT) algorithms ensure continuous tracking.
- Hierarchical Motion Planning: A combination of the Dynamic Window Approach (DWA) for foundational safety and a trained RL value network to sample and select optimal actions for social comfort and compliance.
- Contextual Fine-tuning of RL: Adapting a generic RL model to the specific construction context by creating custom agent behaviors that mimic real-world worker actions (e.g., stop-and-go, pacing, close-range movements). This involves curriculum training and a reward function modification to penalize drastic policy shifts, ensuring gradual and stable model adaptation.
Datasets
The paper does not explicitly mention the use of pre-existing datasets for training the RL model. Instead, it focuses on:
- On-site observations and worker interviews: These qualitative data sources informed the design and the creation of simulated scenarios.
- Custom-designed simulation environment: Used to create experimental scenarios that mimic unique worker behaviors observed on-site.
- Real-world construction site: The primary environment for demonstration and evaluation.
- Lab sandbox environment: Used for controlled quantitative analysis and ablation studies.
Key Insights and Ideas
- Human-centricity is key for construction robotics: Simply automating tasks isn't enough; robots must integrate into existing human workflows and respect human skill.
- "Work companion" is a viable paradigm: Robots can provide significant value by assisting, rather than replacing, skilled human labor, particularly in physically demanding tasks.
- Contextual RL is crucial for dynamic environments: Generic RL models trained in ideal conditions fail in complex, unpredictable settings like construction sites. Fine-tuning RL with context-specific behaviors and environments is essential for safe and effective navigation.
- Perception in cluttered environments is challenging: Distinguishing people from random objects in construction sites requires robust, multi-modal sensing and sophisticated algorithms.
- Balancing 3D mapping accuracy with 2D efficiency: For navigation, a hybrid approach to mapping (3D point cloud projected to 2D) is necessary to capture environmental complexity without prohibitive computational cost.
Key Results
- On-site demonstration: The "work companion rover" successfully performed its intended functions (send, summon, accompany) in three scripted scenarios with human workers. Workers reported feeling safe and comfortable working alongside the robot, finding it a seamless addition to their workflow.
- Lab evaluation:
- The RL-based social navigation layer significantly improved comfortable encounter and delivery rates by approximately 0.4 and 0.15 respectively, compared to a non-RL baseline (DWA).
- The contextualized RL model (DWA+SARL Contextual) achieved a success rate of 1.00, a collision rate of 0.05, and a freeze rate of 0.10, with a navigation time of 25.89 seconds between work zones.
- The contextualized RL model showed improvement over a vanilla RL model, particularly in comfortable encounters and deliveries, demonstrating the importance of context-specific fine-tuning.
Limitations and Future Work
- Limitations:
- The robot's navigation was challenged by a lack of detailed mapping in an extra test case, leading to a localization error.
- Workers expressed a desire for enhanced physical capabilities, such as increased load capacity and the ability to haul carts.
- The current system relies on manual tool/material loading.
- Future Work:
- Expanding the robot's applicability to other construction trades and potential applications.
- Developing autonomous tool/material loading features, possibly with robotic arms.
- Further research into more diverse future robotic support scenarios and approaches.
- Exploring an "ecologically combining" future where human skill and AI/robotics advance together.
Summary updated