# Zero-Shot Task Generalization with Multi-Task Deep Reinforcement Learning

**Authors:** Junhyuk Oh, Satinder Singh, Honglak Lee, Pushmeet Kohli  
**arXiv:** [1706.05064](https://arxiv.org/abs/1706.05064)  
**Date:** 2017-06-15  
**Summary page:** https://www.summarizepaper.com/en/paper/1706.05064v1/

## Key points

- Paper introduces a new problem in reinforcement learning (RL) for executing sequences of instructions after acquiring useful skills for solving subtasks
- Objective is to achieve zero-shot task generalization for previously unseen and longer sequences of instructions
- Proposed objective encourages learning correspondences between similar subtasks through analogies
- Hierarchical architecture with meta controller learns to utilize acquired skills for executing instructions
- Neural architecture in meta controller determines when to update subtask, improving learning efficiency
- Experimental results on stochastic 3D domain demonstrate significance of proposed ideas in achieving generalization to longer and previously unseen instructions
- Research contributes towards developing zero-shot task generalization capabilities in RL by introducing new objectives and architectures

## Summary

The paper titled "Zero-Shot Task Generalization with Multi-Task Deep Reinforcement Learning" introduces a new problem in reinforcement learning (RL) where an agent learns to execute sequences of instructions after acquiring useful skills for solving subtasks. The objective is to achieve zero-shot task generalization, specifically for previously unseen instructions and longer sequences of instructions. To address generalization over unseen instructions, the authors propose a novel objective that encourages learning correspondences between similar subtasks through analogies. For generalization over sequential instructions, they present a hierarchical architecture where a meta controller learns to utilize the acquired skills for executing the instructions. To handle delayed reward, the paper proposes a neural architecture in the meta controller that determines when to update the subtask, thereby improving learning efficiency. Experimental results on a stochastic 3D domain demonstrate the significance of the proposed ideas in achieving generalization to both longer and previously unseen instructions. Overall, this research contributes towards developing zero-shot task generalization capabilities in RL by introducing new objectives and architectures that enable effective execution of instruction sequences after learning subtask-solving skills.