<p>In this paper, a coordinated reinforcement learning (RL)-based distributed optimal locomotion control method is proposed for biped reconfigurable robots (BRRs) under gait task constraints. Based on the joint torque feedback (JTF) technique, a complete distributed dynamic model of BRR subsystem is developed, which incorporates the torso dynamic. To ensure stability during locomotion, the environmental constraints imposed on BRR subsystem are analyzed, and the plantar force distribution is formulated. Then, the reference contact force is derived by utilizing the quadratic programming algorithm. By developing the coordinated learning-based neuro-dynamic programming (NDP) algorithm, the optimal locomotion control of BRRs is achieved, along with comprehensive optimization of environmental contact, plantar force distribution and inter-subsystem constraints. Specifically, a recurrent neural network (RNN) identifier is designed to compensate for inherent model uncertainties, while the actor-critic neural network (NN) approximates the control law and the cost function, which is composed of interaction forces between BRR subsystem, torso, and ground. Then, a distributed near-optimal locomotion control strategy is derived through the online synchronous learning of multiple NNs. Furthermore, Lyapunov-based stability analysis guarantees that the tracking error and NN weight approximate errors of the closed-loop BRR subsystem are uniformly ultimately bounded (UUB). And the experimental validation is conducted on the self-developed claw-type BRRs platform, which demonstrates the effectiveness and reliability of the proposed control method.</p>