A robot path learning and obstacle avoidance system and method combined with deep q-learning

Abstract

The invention discloses a robot path learning and obstacle avoidance system and method combined with deep Q learning. The invention includes an action module, a learning module and an obstacle avoidance module. During path planning, the action module will receive instructions from the learning module and the obstacle avoidance module, and let the robot complete the specified actions according to the instructions. The learning module trains the action selection strategy based on the current state of the robot and the historical data sequence of actions. The obstacle avoidance module implements a random tree search algorithm to guide the robot to obtain a safe path from dangerous situations. After each module is executed, the current state of the robot in the environment will change. The risk of each module is assessed through the scheduling mechanism and the module with the least risk should be activated. The invention solves the problem of large-capacity data storage by using the Q-learning algorithm. Improved the efficiency of the robot when avoiding obstacles.

Description

technical field [0001] The invention belongs to the field of robot control, and in particular relates to a robot path learning and obstacle avoidance method combined with deep Q-learning. Background technique [0002] Exploring unknown environments with mobile robots is a very common problem for robotic applications like rescue and mining. Typically, robots need complex logic about obstacles and a topological map of the environment, aided by information from vision or depth sensors. However, these traditional methods do not have advanced human-brain-like intelligence. The present invention develops a machine learning approach for robots to explore unknown environments using raw sensor input. [0003] To date, research on robot path planning has been extensive, and the literature on modeling and solution methods is extensive. Research related to the realization of threat information can be divided into two categories: static path planning based on prior complete environmen...

AI Technical Summary

Problems solved by technology

Although some self-learning methods can overcome the lack of prior knowledge to some extent, it may suffer from problems caused by over-learning due to various properties of task scenarios.

In contrast, less knowledge of the environment may slow down learning and cause the robot to get stuck in a local optimum

Also, for most learning methods, it is necessary to have a large matrix to hold the computed values

Images