ToribashAI

AI locomotion research using replay datasets, GRU neural networks and Lua runners to learn realistic movement in Toribash.

---

Demo

---

Features

• 🧠 Replay-based locomotion learning
• 🤖 GRU neural network models
• 🏃 Walking and parkour experiments
• 🎯 Automated trajectory evaluation
• ⚡ Lua runner integrated directly into Toribash
• 📊 Evolutionary optimization pipeline

---

Tech Stack

• Python
• PyTorch
• Lua
• Toribash
• JSON
• NumPy

---

Architecture

Replay Dataset
      │
      ▼
Sequence Extraction
      │
      ▼
GRU Training
      │
      ▼
Trajectory Generation
      │
      ▼
Lua Runner
      │
      ▼
Episode Evaluation
      │
      ▼
Evolution Loop

---

Roadmap

• Replay extraction
• Dataset generation
• GRU locomotion model
• Automated Lua runner
• Stable walking
• Dynamic obstacle avoidance
• Parkour navigation
• Reinforcement learning experiments

---

Quick Start

1. Clone the repository

git clone https://github.com/princessnvidia/ToribashAI.git
cd ToribashAI

2. Create a virtual environment

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

3. Install the Lua runner

Copy

data/script/toribash_trajectory_runner_v4_5_reactive.lua

into

Toribash/data/script/

---

4. Start Toribash

/ls toribash_trajectory_runner_v4_5_reactive.lua

---

5. Launch ToribashAI

python scripts/evolution_loop_trajectory_v4_5_reactive.py

---

How it works

The Python process communicates with Toribash through the Lua runner and automatically:

• Generates candidate movements
• Evaluates trajectories
• Scores every episode
• Evolves the population
• Saves the current champion

---

Status

🚧 Active Research Project