Interactive reinforcement learning method for table-balancing robots, and recording media and devices for performing the same

Reinforcement Learning Method for Table Balancing Incorporating Voice Feedback into Rewards

This technology is a robot control system based on Deep Q-Network (DQN) reinforcement learning. It recognizes the state of a table using camera images, performs actions, and optimizes the robot's behavioral policy by performing real-time sentiment analysis on user voice feedback to convert it into reward values.

Existing interactive reinforcement learning methods face challenges such as unnatural interaction due to the use of input devices (mice, remote controls, etc.) and technical limitations requiring specialized domain knowledge for designing reward functions.

This technology establishes an interactive reinforcement learning framework that combines Automatic Speech Recognition (ASR) and sentiment analysis. It automatically converts voice feedback into real-number reward values between -1 and 1 and integrates them into the DQN model's learning reward function, thereby improving learning convergence speed and performance through intuitive human feedback.

‍

‍

Key Features:

• Capturing table state images with a robot camera and transmitting them to a Deep Q-Network (DQN)
• Executing table balancing actions predicted through DQN image analysis
• Evaluating the degree of positivity or negativity of balancing actions based on evaluative voice feedback received from a human
• Calculating reward values based on the positivity/negativity of the voice feedback and outputting balancing actions using a DQN trained over multiple episodes

‍