Curiosity-Driven Robot Vacuum Exploration Study

ISEF Category: Robotics and Intelligent Machines

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

The Hook

A robot vacuum can look smart and still miss half a room. The trick is not just moving fast, it is deciding where to go next. Curiosity-driven systems try to act like a student who keeps checking the parts of the map that still feel new. You can test whether that idea beats a classic exploration rule.

What Is It?

This project studies how a robot can choose places to visit when it does not fully know the space. A normal robot vacuum may follow a simple rule like heading toward the edge of known space, which is called frontier-based exploration. Your topic adds a curiosity signal, built from random network distillation, or RND. RND is a method that gives the robot more reward when it finds places that still look unusual to its internal model.

Think of it like exploring a dark house with a flashlight. Frontier-based exploration walks toward the next doorway it can see. Curiosity-based exploration says, "Keep going where the map still feels unfamiliar." That can help the robot cover tricky spaces, like maze-like floor plans or cluttered tile layouts, more efficiently.

Why This Is a Good Topic

This topic works well because you can measure it with clear numbers, like coverage time, repeated-path rate, and percent of area explored. It connects to real problems in home robotics, search and rescue, and autonomous mapping. You can study an algorithm without needing a human subject or a wet lab. A strong student can learn path planning, mapping, sensor data, and basic machine learning all in one project.

Research Questions

• How does random network distillation affect coverage time compared with frontier-based exploration in maze tiles?
• What is the effect of obstacle density on the advantage of curiosity-driven exploration?
• Does a curiosity-driven policy reduce revisits to already explored tiles more than a frontier-based policy?
• To what extent does sensor noise change the robot's map coverage under each exploration method?
• Which exploration method finds all reachable regions with fewer total turns?
• How does maze complexity change the gap in path efficiency between the two policies?

Basic Materials

• Robot vacuum or small mobile robot platform with programmable control.
• 3D-printed maze tiles or modular floor obstacles.
• Overhead camera or ceiling-mounted phone for top-down recording.
• Tape measure or grid mat for map reference.
• Laptop for logging and analysis.
• Markers or printed labels for zone identification.
• Rechargeable batteries and charger.
• Open-source robotics software stack, if the robot supports it.

Advanced Materials

• Mobile robot base with wheel encoders and lidar or depth camera.
• Onboard computer such as a Raspberry Pi or NVIDIA Jetson.
• ROS-compatible sensor package.
• Calibration targets for camera or lidar alignment.
• 3D-printed maze modules with known geometry.
• External storage for large log files.
• Optional motion-capture or Vicon access for ground-truth trajectory data.
• Access to a workstation for model training and repeated simulation runs.

Software & Tools

• Python: Processes robot logs, computes coverage metrics, and runs statistical tests.
• OpenCV: Tracks robot position from overhead video and extracts path traces.
• ImageJ: Measures explored area from frame captures and map images.
• ROS: Collects sensor data and connects navigation modules on the robot.
• QGIS: Helps compare map coverage on a grid if you convert the maze into spatial layers.

Experiment Steps

1. Define the exact map metric you will use, such as explored area, revisit rate, or time to full coverage.
2. Choose one baseline policy and one curiosity-driven policy so you can compare them fairly.
3. Design maze tiles that create the same level of difficulty across repeated trials.
4. Plan how you will log position, map state, and decision points for each run.
5. Build a scoring method that converts each run into the same set of coverage numbers.
6. Pre-plan your statistical comparison so you can test whether one policy really outperforms the other.

Common Pitfalls

• Using different maze layouts across trials, which makes the policy comparison unfair.
• Tracking the robot with shaky overhead video, which breaks path reconstruction and coverage estimates.
• Comparing raw travel distance without correcting for robot speed, which can hide the real exploration effect.
• Letting the curiosity model train on one maze and testing on a much different maze, which confounds generalization with novelty.
• Ignoring repeated visits to the same tiles, which makes a weak exploration policy look better than it is.

What Makes This Competitive

A strong version of this project goes past a simple head-to-head comparison. You can test several maze types, not just one, and report where curiosity helps or hurts. You can also separate coverage speed from path efficiency, which gives a more honest picture of performance. If you add careful statistics, repeated trials, and a clear failure analysis, the work starts to feel much more like research than a demo.

Project Variations

• Test the same exploration policies in cluttered apartment-like floor plans instead of maze tiles.
• Compare random network distillation with another intrinsic reward method, such as prediction error, on the same robot.
• Analyze how different sensor packages, such as lidar versus camera-only mapping, change the advantage of curiosity-driven exploration.

Learn More

• MIT OpenCourseWare, Introduction to Robotics: Search MIT OpenCourseWare for robotics lectures on navigation and mapping.
• ROS Wiki: Read the navigation and mapping documentation for real robot software workflows.
• Random Network Distillation on arXiv: Search arXiv for the original paper and related intrinsic reward studies.
• PubMed: Search for review articles on autonomous navigation in assistive or home robotics if you want a human-centered angle.
• NASA Open Source Robotics publications: Search NASA technical reports for autonomous exploration and mapping papers.