Deception-Detecting Game Robot

ISEF Category: Robotics and Intelligent Machines

Subcategory: Cognitive Systems  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A tiny pause before an answer can reveal more than the answer itself. Humans leak stress through speech, blinks, and reaction time. A robot that reads those signals could change how it bluffs in a game. That makes this project part psychology, part AI, and part strategy.

What Is It?

This project studies a robot or game agent that watches a human player for signs of deception. Those signs can include micro-pauses in speech, blink rate, and reaction time. The robot then changes how it plays its own bluffing strategy in a Liar's Dice style game.

Think of it like a poker face detector with a brain. A normal game bot only looks at the cards, dice, or rules. Your system also looks at behavior, then asks, “Does this player seem nervous, honest, or ready to bluff?” If the answer changes, the robot changes its next move.

The key idea is not mind reading. The key idea is pattern detection. You are testing whether small human behavior cues predict later game actions well enough to improve win rate against a fixed Bayesian opponent, which is a bot that updates its beliefs with probability math.

Why This Is a Good Topic

This is a strong science fair topic because you can test it with clear numbers, like classification accuracy and win rate. You can start with simple signals from video or audio, then compare them against game outcomes. The project connects to real problems in human-robot interaction, negotiation, security, and adaptive game AI. You can learn data collection, feature extraction, model evaluation, and controlled experimentation without needing a wet lab.

Research Questions

• How does blink rate before a turn predict whether a human player will bluff?
• What is the effect of micro-pauses in speech on the robot's ability to classify deception?
• Does adding reaction time as a feature improve bluff prediction compared with blink rate alone?
• To what extent does adaptive bluffing based on human cues improve win rate against a fixed Bayesian opponent?
• Which combination of cues, blink rate, speech pauses, or reaction time, gives the best deception model?
• How does the robot's strategy change when the opponent's cue patterns become less consistent?

Basic Materials

• Laptop or desktop computer with webcam and microphone.
• Smartphone or webcam for video capture.
• Game dice or digital Liar's Dice simulator.
• External microphone, if the built-in mic is noisy.
• Spreadsheet software for recording outcomes.
• Notebook for trial logs and human observations.
• Consent form for any human participant.

Advanced Materials

• High-frame-rate camera for blink detection.
• Directional microphone for cleaner speech timing data.
• Python-compatible computer with enough memory for video analysis.
• Timestamped game interface or custom game logger.
• OpenCV-compatible video files.
• Audio annotation software.
• Access to a controlled room light setup for repeatable blink capture.
• Optional eye-tracking system for validation.

Software & Tools

• Python: Runs the data pipeline, feature extraction, and model testing.
• OpenCV: Tracks face landmarks and supports blink detection from video.
• Librosa: Extracts speech timing and audio features from recorded turns.
• ImageJ: Checks frame-level visual changes and helps inspect video data.
• Jupyter Notebook: Keeps code, plots, and notes in one place for analysis.

Experiment Steps

1. Define the one deception cue you will test first, such as blink rate, speech pauses, or reaction time.
2. Choose a simple game format that produces repeated bluff decisions you can measure.
3. Plan how you will label each turn, what counts as a bluff, and what counts as a cue.
4. Build a baseline model that predicts bluffing before you add the adaptive robot strategy.
5. Decide how the robot will change its bluffing rule when cue confidence changes.
6. Set up a fair comparison between adaptive play and fixed play, then lock the evaluation metric before testing.

Common Pitfalls

• Recording video with changing lighting, which makes blink detection drift from trial to trial.
• Letting speech recognition replace timing analysis, which can miss micro-pauses that matter more than word choice.
• Training and testing on the same players, which can make the model look better than it really is.
• Defining bluffing too loosely, which creates messy labels and weak results.
• Changing the robot's strategy rules after seeing early scores, which breaks the fairness of the comparison.

What Makes This Competitive

A strong version of this project goes beyond a simple win-rate check. You would separate detection quality from game performance, then show which cues actually carry signal. You could compare several models, test whether the result holds across different players, and use strong statistics instead of a single average score. A more competitive project also studies whether the robot generalizes to new opponents or only learns one person's habits.

Project Variations

• Test whether the same cue model works in a different bluffing game, such as poker-style betting or a simplified liar's sequence game.
• Compare video-only detection with audio-only detection to see which signal type predicts bluffing better.
• Replace the fixed Bayesian opponent with a learning opponent and measure whether the deception model still helps.

Learn More

• MIT OpenCourseWare: Search for introductory machine learning, probabilistic reasoning, and human-computer interaction materials that help with model design.
• NIH PubMed: Search for review articles on deception detection, nonverbal behavior, and reaction time in decision-making.
• NOAA Education: Use communication and data analysis resources for building careful experimental habits.
• NASA Open Data Portal: Explore examples of signal processing, classification, and time-series analysis on real datasets.
• arXiv: Search for recent preprints on deception detection, behavioral cues, and game-playing agents.
• OpenCV Documentation: Read the free guides for face detection, video capture, and landmark tracking.