TLDR: This paper details the first nationwide deployment of Human Activity Recognition (HAR) technology in the on-demand food delivery industry. Utilizing the LIMU-BERT foundation model, the system was successfully adapted for a major Chinese delivery platform, involving 500,000 couriers across 367 cities over two years. The deployment, progressing through feasibility studies and large-scale evaluations, has enabled downstream applications like improved trajectory segmentation, elevation change detection, more accurate Estimated Time of Stop (ETS), and a fairer pricing strategy based on delivery difficulty. Large-scale tests demonstrate significant operational and economic benefits, including an estimated annual saving of 0.44 billion RMB. The paper also shares crucial lessons learned regarding scaling with unlabeled data, large-scale model evaluation, handling diverse devices, and sensor collaboration.
The world of on-demand food delivery is constantly evolving, with companies seeking innovative ways to enhance efficiency and customer satisfaction. A recent research paper highlights a groundbreaking nationwide deployment of Human Activity Recognition (HAR) technology within this industry, showcasing its transformative potential.
The paper, titled Experience Paper: Adopting Activity Recognition in On-demand Food Delivery Business, details how state-of-the-art HAR, specifically the LIMU-BERT foundation model, has been successfully integrated into a major delivery platform in China. This extensive project spanned two years and involved 500,000 couriers across 367 cities, marking the first large-scale commercial adoption of HAR in this sector.
The Challenge of Real-World Activity Recognition
Implementing HAR in a dynamic environment like food delivery presents unique hurdles. Couriers use a vast array of smartphone models (over 1,000), leading to significant data variations. Unlike controlled lab settings where devices are fixed, couriers constantly interact with their phones, changing placement and introducing noise into sensor data. Furthermore, collecting vast amounts of labeled data for training is incredibly expensive and time-consuming.
LIMU-BERT: A Smart Solution
To overcome these challenges, the researchers adapted LIMU-BERT, a powerful sensor foundation model. This model excels because it can learn effectively from massive amounts of unlabeled sensor data, which is readily available from delivery apps. It works in two phases: first, it’s pre-trained on billions of unlabeled data samples to understand general sensor patterns; then, it’s fine-tuned with a smaller, labeled dataset for specific activity recognition tasks like identifying if a courier is walking, riding, or still. Its lightweight design also ensures it runs efficiently on smartphones without draining battery or increasing app size significantly.
A Phased Deployment Across China
The deployment unfolded in three key phases:
-
Phase I (Feasibility Study): Began in January 2022 in Yangzhou City. This phase focused on customizing LIMU-BERT for delivery scenarios, including optimizing data sampling rates and window sizes. Initial tests showed LIMU-BERT significantly outperformed traditional models in recognizing activities like ‘still,’ ‘walking,’ and ‘riding.’
-
Phase II (Large-Scale Evaluation): From July to December 2022, LIMU-BERT was pre-trained using billions of unlabeled data samples from 60,000 couriers nationwide. To evaluate its performance at scale without extensive manual labeling, a rule-based approach was used, classifying activities as ‘riding’ or ‘non-riding’ based on GPS speed and indoor/outdoor detection. The model achieved over 90% accuracy in this large-scale assessment.
-
Phase III (Online Deployment): Starting in June 2023 and continuing to the present, the HAR models have been gradually rolled out to over 500,000 couriers across 367 cities. The system now processes approximately 7.5 billion predictions daily, demonstrating its robustness and scalability in real-world operations.
Transformative Applications and Business Benefits
The adoption of HAR has enabled several crucial applications, leading to significant operational and economic advantages:
-
Trajectory Segmentation and Navigation: By accurately identifying courier activities (e.g., riding to walking), the system can pinpoint key moments like arrival at a restaurant or customer location. This improves navigation recommendations, especially for complex drop-off points, enhancing courier efficiency.
-
Elevation Change Detection: LIMU-BERT can detect vertical movements (like using elevators or stairs) even on phones without barometer sensors. This provides more detailed insights into courier behavior, crucial for understanding delivery challenges within multi-story buildings.
-
Estimated Time of Stop (ETS): More precise activity recognition leads to better predictions of how long a courier will stop at a location. This refinement reduces the mean absolute error (MAE) of ETS by 1.8 seconds per order overall, significantly improving Estimated Time of Arrival (ETA) accuracy and customer satisfaction.
-
Difficulty Analysis and Pricing Strategy: The system can identify “difficult” delivery locations (e.g., restaurants on high floors, complex customer areas) based on activity patterns. This allows for a fairer pricing strategy, increasing delivery fees for challenging orders and decreasing them for easier ones. This dynamic pricing is estimated to save the platform approximately 0.44 billion RMB annually.
Also Read:
- Digital Twins in Industrial Maintenance: A Comprehensive Review of Predictive Strategies and Future Directions
- Empowering AI to Recognize Its Own Limits in Complex Reasoning
Key Lessons Learned from Deployment
The large-scale deployment offered valuable insights:
-
The Power of Unlabeled Data: Leveraging vast amounts of unlabeled sensor data is critical for training robust models that generalize well in diverse real-world scenarios.
-
Scalable Evaluation: For large-scale deployments, rule-based labeling using other sensors (like GPS) provides a cost-effective and scalable way to evaluate model performance, complementing expensive manual annotations.
-
Handling Device Diversity: It’s essential to adapt models for different smartphone capabilities. For instance, an accelerometer-only LIMU-BERT was developed for devices lacking gyroscopes, increasing courier coverage from 89% to 99%.
-
Sensor and Device Collaboration: Combining data from various sensors (GPS, light, magnetometer, barometer) and even leveraging data from different device types (e.g., iOS barometers to train Android models) enhances overall system robustness and applicability.
-
Practical Optimizations: Techniques like dynamic loading of software development kit (SDK) components were crucial to minimize app size, addressing a key commercial constraint.
This pioneering work demonstrates the immense practical value of human activity recognition in optimizing operations and achieving substantial economic benefits in the fast-paced on-demand food delivery industry. It paves the way for future advancements in mobile computing and AI applications in real-world commercial settings.
