2.3. Data Generation

Sensorimotor data is generated from a mixture of behaviors: motor babbling and goal-directed movements targeting the object. Motor babbling assigns random velocities to the motor joints of the arms (Δm_i for i ∈ 0, 1, 2 for the bottom agent and i ∈ 3, 4, 5 for the top one) at different moments. We need independent motor babbling to not cause an artificial correlation between the agents due to the synchronous randomly generated behavior. We sustain, for different time steps, the assigned motor torque to each joint m_i. We need to sustain the torque applied to generate a consistent movement with a consistent effect during a certain time. Because of the characteristics of the setup, we needed to set certain rules on top of the independent random signals sent to the motor joints to generate a rich and representative dataset. For example, when the object reached an end-point under one of the two boxes, the object was reset in the middle again, and the two arms were positioned at their initial positions. As an example, in Figure 1B, we show the time series of s₁ and s₆ SIPs before and during an object contact.

Generating the behavior of agents in this way, we acquire a dataset of all SIP data streams (position, velocities, accelerations, and haptic contacts) and we are thus ready to apply the corresponding statistical tools to determine the nature of every SIP.