The control problem with the humanoid model has been a challenging question. To develop a suitable control system for the humanoid model to complete human tasks, a recent approach is imitation learning. Generative Adversarial Imitation Learning (GAIL) is the state-of-the-art imitation learning framework available. With the Imitation from Observation (IfO) variant, the GAIL framework can learn to imitate motion from the mocap data for the humanoid model. Since the GAIL framework requires the target motion to be recorded with the expensive mocap system, it is hard for GAIL to control the humanoid model to imitate new motion that is not recorded with the mocap system. Most of the time the GAIL framework can only imitate one single target motion, which makes it impossible to develop a cont...[ Read more ]

The control problem with the humanoid model has been a challenging question. To develop a suitable control system for the humanoid model to complete human tasks, a recent approach is imitation learning. Generative Adversarial Imitation Learning (GAIL) is the state-of-the-art imitation learning framework available. With the Imitation from Observation (IfO) variant, the GAIL framework can learn to imitate motion from the mocap data for the humanoid model. Since the GAIL framework requires the target motion to be recorded with the expensive mocap system, it is hard for GAIL to control the humanoid model to imitate new motion that is not recorded with the mocap system. Most of the time the GAIL framework can only imitate one single target motion, which makes it impossible to develop a control system that can complete human tasks, since those tasks may require the skill of multiple target motions. Inspired by how a human learns a new motion, the framework proposed in this thesis uses the primary motion which is recorded in mocap data and learns to imitate the unstable target motion taped in video clips with the assistance of the learned skill of the primary motion. This framework also introduces the idea of the external control signal such that the framework can change to imitate different target motions in real-time. With the aid of the primary motion, the proposed framework can deliver a stable and human-like motion that imitates multiple target motions taped in inaccurate video clips. With this framework, the humanoid model can learn to imitate different target motions from low-cost sources and can interchange the imitating target motion in real-time to complete the human task.