Monte Carlo Tree Search with Tensor Factorization

This repository provides the official implementation of Tensor Train Tree Search (TTTS), a novel algorithm that integrates Monte Carlo Tree Search (MCTS) with Tensor Train (TT) factorization. TTTS efficiently handles complex decision-making problems in robotics by combining global search, local refinement, and a low-rank representation of the decision tree.

TTTS enables scalable and parallelizable planning across a variety of tasks, including nonlinear dynamics, hybrid discrete-continuous optimization, non-convex constraints, and multi-modal solution discovery.

---

🔬 Features

• Combines MCTS with tensor factorization to reduce combinatorial complexity.
• Supports non-linear kinematics, non-convex constraints, mixed-integer programming and complex contact-rich dynamics.
• seamless integration with black-box simulators (e.g., Genesis).
• Supports GPU-based parallel computation.

---

📁 Experiments

In the examples folder, you can find the following examples:

1. Toy Examples

• Continuous Non-Convex Optimization
• Mixed-Integer Optimization

If you are interested in a comparison between Tensor Trains (TTs) and neural networks (NNs) as function approximators, check out tt_vs_nn.ipynb.

2. Inverse Kinematics

• Solves collision-free IK for multi-joint manipulators

• Demonstrates TTTS’s ability to find multiple solutions

3. Motion Planning Around Obstacles

• 3-joint Manipulator: Reaches a narrow target via long-horizon planning
• 7-joint Panda Arm: Moves between shelves while avoiding collisions

4. Bimanual Whole-Body Manipulation

• Real-world task using whole-body contact
• Solves complex control with real-time MPC using TTTS

---

🚀 Getting Started

pip install -r requirements.txt

📬 Contact

For questions or contributions, please contact:

Teng Xue [teng.xue@idiap.ch]