Abstract
While advancements in computer vision and natural language processing have significantly enhanced robot perception, the sense of touch remains a frontier in social robotics. This paper introduces a comprehensive pipeline to enable natural tactile interaction for robot dogs using high-density, large-format flexible pressure sensors. By treating human-robot interaction as a translation task, the system maps 81 distinct human gestures to lifelike canine responses. Our results show a 98.7% gesture recognition accuracy and demonstrate that tactile interaction is significantly more effective than voice commands in fostering emotional bonds and reducing user anxiety.
Introduction
The sense of touch is fundamental to building affective connections between humans and animals. However, most current robotic tactile systems are restricted to specific end-effectors or utilize low-density sensors that cannot capture the nuance of complex social interactions like stroking, scratching, or hugging. For social robots, particularly quadrupedal "robot dogs," this lack of high-fidelity tactile feedback prevents them from being perceived as convincing companions. This research addresses this gap by developing a whole-body "electronic skin" and an intelligent framework to interpret and respond to human touch naturally.
Methods: Hardware and Software Pipeline
The proposed system integrates a hardware-to-software pipeline designed to capture and process high-resolution tactile data in real-time.
Large-Format Tactile Skin
The researchers outfitted a Xiaomi CyberDog with 13 piezoresistive film sensors. This "skin" covers the head, back, hips, and legs. The configuration includes six 32×32 sensor arrays (each providing 1024 sensing points), three 13×16 arrays, and four 1×1 sensors for smaller areas like the chin. Data is sampled at 10Hz through custom reading boards, producing "tactile images" that represent the pressure distribution over the robot's body.
Gesture Recognition and Action Prediction
The software pipeline processes these tactile images in two main stages:
- Perception: A modified ResNet34 architecture classifies sequences of 20 tactile frames (representing a 2-second temporal window) into one of 81 gesture categories. These categories represent combinations of 13 basic gestures (e.g., massage, tickle, cradle) and 11 body parts.
- Behavior Mapping: Instead of using simple "if-this-then-that" rules, the system employs a Transformer model to "translate" human gestures into sequences of dog actions. By treating HRI as a machine translation task, the robot considers the context of previous interactions, leading to more varied and less robotic responses.
Experimental Results
The system was evaluated through both technical benchmarks and a human-robot interaction study involving over 300 participants.
Technical Performance
The ResNet-based gesture recognition model achieved an impressive 98.7% accuracy across the 81 defined gesture categories. Furthermore, the Transformer-based action prediction model reached a 1-gram BLEU score of 0.87, indicating that the robot's generated responses closely align with the natural behaviors observed in biological dogs during similar interactions.
Psychological Impact
A comparative study between tactile and voice-based interaction revealed profound differences in user experience. Participants using touch reported a stronger emotional connection, viewing the robot more as a pet than a machine. Tactile interaction was found to be significantly more effective at alleviating anxiety and stimulating excitement. Notably, the average interaction time for the tactile group (12.5 minutes) was more than double that of the voice-command group (5.5 minutes), suggesting much higher engagement levels.
Discussion and Conclusion
This work demonstrates that high-density tactile sensing is a viable and powerful modality for social robotics. By framing interaction as a translation problem, robots can exhibit more nuanced and context-aware behaviors that mirror biological counterparts.
Despite these results, hardware constraints remain; for instance, the rigid neck and tail of the current robot limit its expressive range. Future work will focus on multi-modal integration—combining touch with vision and voice—to create a more holistic interaction experience. Additionally, moving toward an end-to-end learning model could further improve the efficiency and interpretability of the tactile processing pipeline.
Summary updated