Robot Maze Navigation with Vision and Proximity Sensing

Teleoperated data collection, map reconstruction, and imitation-learning policies for autonomous maze navigation.

Overview

This repository presents a complete robot maze-navigation pipeline built around teleoperation, sensor-based mapping, supervised policy learning, and autonomous execution. The project combines RGB vision, proximity-sensor measurements, odometry, and reconstructed maze structure to support end-to-end driving policies.

Two policy families are included:

• a single-frame TensorFlow/Keras policy
• a temporal PyTorch policy

The repository also includes reconstructed maze maps, evaluation artifacts, demo videos, and the best checkpoints for both approaches.

Highlights

• End-to-end workflow from teleoperation to autonomous policy execution
• Vision and proximity sensing used together for navigation
• Maze reconstruction from logged robot motion and sensor data
• Two learning pipelines:
  • single-frame TensorFlow/Keras baseline
  • temporal PyTorch policy
• Curated evaluation metrics, confusion matrices, ROC/PR curves, and demo videos

Results

• TensorFlow/Keras single-frame policy
  • accuracy: 97.99%
  • evaluation set size: 4,716
• Temporal PyTorch policy
  • accuracy: 98.83%
  • evaluation set size: 20,457
  • expected calibration error: ≈ 0.035

Repository structure

vision-proximity-maze-navigation/
├── scripts/
│   ├── teleop/
│   ├── mapping/
│   ├── keras/
│   └── pytorch/
├── artifacts/
│   ├── maps/
│   ├── metrics/
│   │   ├── keras/
│   │   └── pytorch/
│   └── media/
├── checkpoints/
│   ├── keras/
│   └── pytorch/
├── data/
├── requirements-common.txt
├── requirements-keras.txt
├── requirements-pytorch.txt
└── README.md