CARDinality: Interactive Card-shaped Robots with Locomotion and Haptics using Vibration

Aditya Retnanto*, Emilie Faracci*, Anup Sathya*, Yu-Kai (Kain) Hung, Ken Nakagaki
University of Chicago & National Taiwan University · UIST '24, Pittsburgh, PA · *equal contribution

ACM DL · arXiv · AxLab project page

About

CARDinality is a new approach to interactive robots that borrows the form factor of a playing card. Each robot is a thin, flat device that uses vibration both to slide across a surface in any direction and to give haptic feedback. With on-device sensing and wireless control, the robots are light, portable, and scalable, and they open up tangible interactions for card games, learning tools, and everyday objects.

My contributions: fabrication (3D printing and soldering), locomotion training, the sensing pipeline and algorithms, Bluetooth networking, and experiment design.

Hardware

A 0.3 mm semi-flexible PCB measuring 56 × 89 mm, close to a standard 63 × 89 mm playing card. The robot is 1.3 mm at its thinnest point and 4.7 mm at its thickest.
Four surface-mount eccentric rotating mass (ERM) vibration motors, each driven in both directions by two dual-channel motor drivers.
A Seeed Studio XIAO nRF52840 Sense for BLE, an onboard 6-axis IMU, and USB-C charging.
A 180 mAh LiPo battery that lasts about 35 minutes at full intensity and more than two days in standby.

Vibration-based Omni-directional Locomotion

Turning four motor intensities into a direction of travel is hard to model, so each robot learns its own motion patterns. In a camera-tracked "robot school" arena with ArUco markers, a grid search runs 625 motor-intensity combinations and measures the resulting motion. The best patterns cover 36 headings at 10° steps, plus clockwise and counter-clockwise rotation. All 76 patterns are stored in firmware, so a single 1-byte BLE command can trigger any of them.

Sensing

State: face-up or face-down on a table, or held in the hand, detected from accelerometer orientation.
Gestures: taps and slides on the table, and shakes and flicks in the hand, classified with logistic regression.
Surface texture: the robot vibrates one of its own motors and reads the response with its IMU through an FFT. It identifies the surface it is on with 74–95% accuracy across table, rubber, felt, and card-mat surfaces.

A Flask-based control interface handles omni-directional driving, haptic presets, and camera-based closed-loop targeting. One laptop has run up to 10 robots at the same time.

Applications

Augmented card games: monster duels in which cards buzz, move to stage collisions, and rotate to show who won, plus a haptic matching game.
Actuated flashcards: vocabulary cards that prompt the learner and sort themselves into "known" and "review" zones.
Everyday cards: a credit card that confirms payments through haptics and slides away during impulse purchases, and a key card that reports lock status and can slide under a door.

Limitations & Future Work

Motion patterns depend on the individual device and the surface. A model transferred from one robot to another performs poorly, and sleeves or textured bases call for retraining. Next steps include onboard localization in place of the external camera, thinner replacements for the rigid components that cover about 22% of the board, and patterns that generalize across devices and surfaces.