SpaceTools: Tool-Augmented Spatial Reasoning via Double Interactive RL

Official repository for SpaceTools and the Toolshed system.
🛠️ Code coming soon — we are preparing a full release including the Toolshed infrastructure, DIRL training pipeline, and evaluation scripts.

---

📌 Overview

SpaceTools is a framework that empowers VLMs with vision tools and robotic tools for spatial reasoning and real-world manipulation.
It introduces Double Interactive Reinforcement Learning (DIRL), a two-phase training pipeline that enables effective multi-tool coordination.

The code release will include both:

🔧 1. Toolshed (System Release)

A scalable infrastructure for deploying compute-heavy tools during both training and inference:

• Isolated environments for each tool
• Decoupled resource scaling
• Async parallel workers per tool
• Support for heavy tools (segmentation, pointing, depth, 3D box, grasp prediction)

🧠 2. SpaceTools (Model + Training Release)

• DIRL training pipeline
• SFT + RL dataset
• Tool-augmented inference
• Spatial benchmark evaluation

For project details and demos:
👉 Project Page: https://spacetools.github.io/
👉 Paper: https://arxiv.org/pdf/2512.04069

---

📦 Installation (coming soon)

We will provide detailed setup instructions, including:

• Recommended environment (conda / pip)
• CUDA / PyTorch version requirements
• Setup for Toolshed workers and servers
• Integration of Toolshed for interactive RL
• Integration of Toolshed for zero-shot frontier model reasoning
• Supervise fine-tuning for tool use
• Dependencies for VLMs, RL, SFT, and each tool backend

# Placeholder – installation instructions coming soon
git clone https://github.com/spacetools/spacetools.git
cd spacetools