# Combining Neural Networks and Tree Search for Task and Motion Planning in Challenging Environments

**Authors:** Chris Paxton, Vasumathi Raman, Gregory D. Hager, Marin Kobilarov  
**arXiv:** [1703.07887](https://arxiv.org/abs/1703.07887)  
**Date:** 2017-03-22  
**Summary page:** https://www.summarizepaper.com/en/paper/1703.07887v1/

## Key points

- The paper explores task and motion planning in complex dynamic environments.
- The authors propose a methodology that combines reinforcement learning and deep neural networks.
- The objective is to generate task plans based on Linear Temporal Logic (LTL) constraints and a reward function.
- Monte Carlo Tree Search is integrated with hierarchical neural net control policies for efficient exploration.
- The paper investigates the ability of neural networks to learn LTL constraints and control policies.
- The approach is demonstrated in a simulated autonomous driving scenario.
- Overall, the research contributes to task and motion planning by leveraging deep neural networks and reinforcement learning techniques.

## Summary

The paper titled "Combining Neural Networks and Tree Search for Task and Motion Planning in Challenging Environments" explores the problem of task and motion planning in complex dynamic environments. The authors propose a methodology that utilizes reinforcement learning and deep neural networks to learn low-level control policies and task-level option policies. The objective is to generate task plans based on a set of Linear Temporal Logic (LTL) constraints and a reward function. One of the main challenges in this setting is the need to explore future scenarios in a complex and interactive environment. To address this, the authors integrate Monte Carlo Tree Search with hierarchical neural net control policies trained on expressive LTL specifications. This combination allows for efficient exploration of possible outcomes. The paper investigates the ability of neural networks to learn both LTL constraints and control policies. The authors demonstrate their approach in a simulated autonomous driving scenario, where a vehicle must navigate through traffic, avoid collisions, and navigate intersections while adhering to predefined rules of the road. Overall, this research contributes to the field of task and motion planning by leveraging deep neural networks and reinforcement learning techniques to generate effective task plans in complex environments. The integration of Monte Carlo Tree Search enhances exploration capabilities, enabling better decision-making in dynamic scenarios.