Figure 2.

The modular behavioral environment (MoBeE) architecture. MoBeE implements low-level control and enforces all necessary constraints to keep the robot safe and operational in real time, such that the curious RL agent (left) is able to experiment with arbitrary control commands. A kinematic/geometric model of the iCub humanoid robot (top) is driven by streaming motor encoder positions from the hardware (right). The model computes fictitious constraint forces, which repel the robot from collisions, joint limits, and other infeasibilities. These forces, f_i(t) in Equation (1), are passed to the controller (middle), which computes the attractor dynamics that governs the actual movement of the robot.