Fig. 8
From: Deep reinforcement learning for modeling human locomotion control in neuromechanical simulation
OpenSim-RL environment for the NeurIPS 2019: Learn to Move - Walk Around competition. a A neuromechanical simulation environment is designed for a typical RL framework (Fig. 5). The environment took an action as input, simulated a musculoskeletal model for one time-step, and provided the resulting reward and observation. The action was excitation signals for the 22 muscles. The reward was designed so that solutions following target velocities with minimum muscle effort would achieve high total rewards. The observation consisted of a target velocity map and information on the body state. b The environment included a musculoskeletal model that represents the human body. Each leg consisted of four rotational joints and 11 muscles. (HAB: hip abductor; HAD: hip adductor; HFL: hip flexor; GLU: glutei, hip extensor; HAM: hamstring, biarticular hip extensor and knee flexor; RF: rectus femoris, biarticular hip flexor and knee extensor; VAS: vastii, knee extensor; BFSH: short head of biceps femoris, knee flexor; GAS: gastrocnemius, biarticular knee flexor and ankle extensor; SOL: soleus, ankle extensor; TA: tibialis anterior, ankle flexor). c The simulation environment provided a real-time visualization of the simulation to users. The global map of target velocities is shown at the top-left. The bottom-left shows its local map, which is part of the input to the controller. The right visualizes the motion of the musculoskeletal model