双足机器人步态控制的深度强化学习方法

Deep reinforcement learning method for biped robot gait control

针对双足机器人行走过程中的步态稳定控制问题,提出一种改进深度Q网络的深度强化学习方法。首先,将深度Q网络算法与确定性策略梯度相结合,提出用修正Double-Q网络优化操作—评论网络的评论网络,给出一种改进的深度Q网络;然后,建立双足机器人连杆模型,在常规的平整路面上将改进的深度Q网络用于作为智能体的双足机器人进行步态控制训练。MATLAB仿真结果表明,与深度Q网络和深度确定性策略梯度算法相比,所提算法有更好的训练速度且其回报曲线具有良好的平滑性。在CPU训练下,经过20 h左右深度强化学习能够完成智能体训练。双足机器人在较小的力矩和长距离下能够稳定快步行走。

Aiming at the stable control of gait during biped robot walking,a deep reinforcement learning method with improved Deep Q-Network(DQN)was proposed.By combining DQN algorithm with a deterministic strategy gradient,an improved DQN learning network was proposed to replace the critic network of actor-critic network with a clipped Double-Q network.A link model of biped robot was established,and the proposed network was used for biped robots gait control training as agents in a conventional flat road environment.MATLAB simulation results showed that compared with DQN and Deep Deterministic Policy Gradient(DDPG)algorithms,the proposed algorithm had a better training speed and its average reward curve had a good smoothness.Under the CPU training conditions,the agent training could be completed after about 20 hours of deep reinforcement learning.The biped robot could achieve stable and fast walking(average speed about 0.5m/s)under the conditions of small torque and long distance(about 5 meters).