Squat and tuck jump maneuver for single-legged robot with an active toe joint using model-free deep reinforcement learning

Nowadays, legged robots are becoming increasingly popular. They are well-suited to rescue operations or factory inspections because of their ability to serve in uneven terrains. Single-legged robots are sometimes used due to their high maneuverability and simple structures that can be extended to more complex robots later. Jumping is one of the locomotion types of legged robots. A jump can be regarded as a precursor to running, in addition to being able to pass through obstacles that they would typically encounter. In this study, a single-legged robot with active toe joints is designed based on human biomechanics. The goal is to train an agent to do squats and tuck jumps directly from data without path planning. Therefore, reinforcement learning is used in the low-level control layer. Next, a study is conducted to determine how the complexity of reward functions affects robot behavior. Four different model-free reinforcement learning, algorithms are used to train agents on deterministic and stochastic policies. A study revealed that the agent learned how to use its toe actively to achieve more height. In conclusion, the jumping heights of these algorithms are compared according to their averages and variances.