Note: This is a remake of a 2020 project, rebuilt from scratch after the original code was lost — metrics may be a little bit different from the research report.

Darkchess Robot

🤖 Demo

💡 Getting Started

Setup

Use pixi to set up the development environment:

pixi shell

Run Darkchess Robot System with Specific Mode

# Run Robot Server along with Website Monitor
python app.py --robot

# Run API Server of 'brain' & 'eye' Only
python app.py --api

Go config.py to configure app settings before running server.

🚀 Features

Brain - AI Engine for Darkchess Board Game

• Search-based Agents
  • Random: Just Random.
  • Min-Max: 60% win rate versus human.
  • Alpha-Beta: Faster version of Min-Max.
• Learning-based Agents
  • Q-Learning (QL): Tabular Q-learning with a discrete Q-table.
    • Metrics and Models: Normal mode (8×4), Small mode (3×4).
  • Q-Learning with MCTS (QL-MCTS): QL combined with PUCT MCTS for lookahead search.
    • Metrics and Models: Normal mode (8×4), Small mode (3×4).
  • Deep Reinforcement Learning (DRL): Q-learning with a neural network (DQN) replacing the Q-table.
    • Metrics and Models: Normal mode (8×4), Small mode (3×4).
  • Deep Reinforcement Learning with MCTS (DRL-MCTS): DRL combined with PUCT MCTS for lookahead search.
    • Metrics and Models: Normal mode (8×4), Small mode (3×4).

Overall: DRL-MCTS ≈ DRL ≥ Alpha-Beta = Min-Max > QL-MCTS > QL > Random.
For small mode (3x4), read here for detail.
For Brain training/testing script, refer to brain_train.ipynb.

Eye – Real-World Detection and Recognition of Darkchess Board States

• Model Architecture: VGGNet-based darkchess board recognition
• Training Accuracy: 99.9%
• Training Loss: 7.2336e-06
• Real-world Test Success Rate: 98.9%

For Eye training/testing script, refer to eye_train.ipynb.
Detailed model information is available on Huggingface.

Arm – Robotic Arm Control for Real-World Darkchess Applications

Our Third-Generation Catcher model with robotic arm that doing Chess-Flipping actions in Real-World Darkchess game.

Explore the Hardware and Firmware for more details.

AIoT – Darkchess Robot System Operable via WiFi

• API server collects training data from both Eye and Brain features.
• Just a single button press starts gameplay via WiFi remote control.

For API specifications, please see documentation.

🌟 Awards

• Gold Medal (Int'l Top 50/408) - 7th Kaohsiung International Invention & Design EXPO (2020)
• Merit (Nat'l Top 20/661) - 19th Macronix Science Awards (2020)
• Second Place (Nat'l 2/200up) - 12th i-ONE NARLabs Instrument Technology Innovation Competition (2020)
• Second Place in Engineering(I) (Nat'l 2/151) - 60th National Primary and High School Science Fair (2020)
• First Place in Engineering(I) (Reg) - 53rd Taipei Primary and High School Science Fair (2020)
• First Place in Engineering(I) (H.S.) - Taipei Municipal Neihu Vocational High School Independent Study (2020)

®️ Patent

• Invention Patent - DARK CHESS ROBOT (Dec 1, 2021) - TWI748780B
• Utility Model Patent - Robot arm gripper (Feb 21, 2021) - TWM608235U

Please visit here and search for the patent code as mentioned above.