基于深度/单目融合视觉及强化学习的机器人定位棋局与行棋策略

中国象棋对弈机器人系统实现的关键包括棋局识别定位和自主行棋策略.:首先,针对棋局识别与定位问题,提出一种基于单目相机与深度相机视觉融合的棋局识别定位方法.:该方法利用立体棋子三维特征获取棋子位置,与二维图像识别结果融合计算定位,以提高棋子的识别定位精度.:其次,针对行棋策略问题,提出一种基于深度神经网络与蒙特卡洛树搜索的决策方法.:该方法利用具有终局特征判断的蒙特卡洛树进行搜索,使用优化的随机行棋策略指导模拟行棋,训练具有多尺度及残差结构的策略价值网络模型.:最后,通过自对弈获取训练数据,通过智能体对抗验证、更新模型参数.:实验表明,相较于单目视觉识别,所提出方法具有更高的精确度和稳定性,识别率达到97%;相较于基准剪枝搜索算法,所提出方法对弈时最多赢得82%的对局,且所需运算时间缩短41%.

Chess positioning and playing strategy of robot based on integrated depth/mono vision and reinforcement learning

The key to the realization of the Chinese chess system lies in the board recognition and chess strategy. Firstly, for the problem of chessboard recognition, a method based on the fusion of mono vision and depth vision is proposed. This method designs a chess piece grid recognition network, uses the three-dimensional characteristics of chess pieces to convert the depth image into a chessboard grid, and integrates the chess piece coordinates with the chessboard grid to effectively improve the recognition accuracy of the chessboard. Secondly, aiming at the problem of the chess strategy, a method based on the deep neural network and Monte-Carlo tree search is proposed. This method uses the improved random search strategy with end-game feature judgment to guide the simulation of chess, which trains a policy and value network with residual structure. Finally, the training data is obtained through self-playing, and the parameters are updated and verified through the agent confrontation. Experiments show that compared with mono-only visual recognition, this method has higher accuracy and stability, and the recognition rate reaches 97%. Compared with the pruning search algorithm baseline, this method wins 82% of the games, and the computing time is reduced by 41%.