Authors: Shaopeng Zhai, Jie Wang, Tianyi Zhang, Fuxian Huang, Qi Zhang, Ming Zhou, Jing Hou, Yu Qiao, Yu Liu

Published on: December 12, 2023

Impact Score: 8.3

Arxiv code: Arxiv:2401.00006

Summary

• What is new: OpenContra co-trains pre-trained language models and goal-conditioned reinforcement learning to create an agent that understands and executes any human instruction.
• Why this is important: Existing agents struggle with context-specific interactions and efficient exploration.
• What the research proposes: A co-training framework that combines the strengths of language models and reinforcement learning to understand and execute arbitrary goals.
• Results: The agent successfully comprehends and completes a wide range of tasks with a high completion ratio in a complex game environment.

Technical Details

Technological frameworks used: OpenContra

Models used: Pre-trained language models, Goal-conditioned reinforcement learning

Data used: Contra game environment

Potential Impact

Gaming, robotics, and automation industries

Want to implement this idea in a business?

We have generated a startup concept here: AdaptiBot.