An acceleration method for deep reinforcement learning of simulated robots

Abstract

The invention belongs to the field of robots, and discloses an acceleration method for deep reinforcement learning of a simulated robot, with the purpose of accelerating the learning process so as to reduce the time cost of research, debugging and deployment of robot deep reinforcement learning. The technical scheme of the present invention is: select one node as a learning node, and other nodes as environment nodes; each environment node processes the interaction details with a robot emulator instance, and provides a unified environment interaction message interface; the learning node adopts the form of frame simulation Interact with each environment node through a message interface, and collect learning data from multiple environments at the same time, thereby accelerating reinforcement learning. The present invention decouples learning algorithm development and simulation interaction details through the abstraction of environment nodes while adapting to various robot simulators, and message communication allows each environment node and simulator instance to be deployed in a distributed computing environment, which is easy to deploy , Scalable advantages.

Description

technical field [0001] The invention belongs to the field of robots, and relates to an acceleration method for deep reinforcement learning of robots in a simulation environment, and can be applied to robot control tasks such as obstacle avoidance, navigation, formation, and multi-robot collaboration of intelligent robots. Background technique [0002] Reinforcement learning is one of the important technologies used in the field of robotics. Through reinforcement learning, robots can learn a set of action strategies to complete tasks independently through continuous attempts. This self-learning ability plays an important role in complex scenarios where it is difficult to manually design action strategies. significance. [0003] Reinforcement learning is used to solve sequential decision-making problems. The learner (that is, the agent) tries to make an action according to the action strategy combined with the current environment state (the initial strategy is usually a random...

AI Technical Summary

Problems solved by technology

For simulation-based learning, on the one hand, the simulation environment is required to be as realistic as possible, so that the learning results can be more easily transferred to reality, but on the other hand, the more realistic the simulation environment, the greater the computational cost. The cost of training time, the evolution of the simulation environment has become a bottleneck that limits the speed of deep reinforcement learning for robots

Images