Imitation Learning by Estimating Intention of Others

By imitation of others, we can obtain the "theory of mind" of one another, which improves communication. Therefore, a talent for imitation learning is essential. Generally, it is difficult to imitate others as the dynamics and kinematics are different for the imitator and the demonstrator. For example, when children imitate its parents, it must adapt the observed behaviour to its own, smaller body.

We propose an imitation learning algorithm where the imitator (learner) estimates the intention of the demonstrator. The agent can imitate even if there are differences in dynamics and kinematics by acting according to the estimated intention. We tested the algorithm for a cart-pole swingup task where the body of the imitator is bigger than for the demonstrator. Fig. 1 shows the trajectory of the demonstrator, and fig. 2 shows the trajectory of the imitator. After several observations, the imitator could succesfully swing up. The control law of the imitator uses only the estimated intention of the demonstrator, leaving the dynamics unaffected by the difference in size.