Fig. 1. The configuration of RNNPB in learning phase (a) and interaction phase (b).

Fig. 2. The rolling a ball behavior (a) and the lifting up a ball behavior by QRIO.

Fig. 3. Snapshots of the rolling a ball behavior (a) and the lifting up a ball behavior (b).

Fig. 4. Dynamic generation and switching of two learned behaviors in the ball handling task. Plot in the top row represents the measured position of the ball. Plot in the second and the third row represent the ball position predicted and the robot joint angles generated by the RNNPB. Plot in the bottom row represents the parametric bias of the RNNPB.

Fig. 5. Prediction error distribution in the PB space. Only the area where the prediction error is less than 0.00011 is shown. The right half contour indicates the area to realize ball rolling behavior and the left half contour indicates the area to realize the ball lifting up behavior.

Fig. 6. Ball rolling behavior according to constant PBs by the RNNPB, which has learned two different types of ball handling behaviors. Each graph labeled with (a)–(c) indicates the results that were acquired from the constant PB selected from the PB space where the prediction error corresponds with rolling, intermediate and lifting up behavior consequently. Plots in the top and second row represent the calculated target velocity of ball movement and the calculated velocity based on prediction by the RNNPB for comparing the performance of one-step prediction by the RNNPB. Plots in the bottom row represents the joint angles generated by the RNNPB.

Fig. 7. Ball lifting up behavior according to constant PBs. The labels below the graphs have the same meaning as Fig. 6.

Fig. 8. Snapshots of the behavior switching guided by the human user.

Fig. 9. Guided behavior switching in the block building task. Plot in the top row represents the measured position of three blocks. Plot in the second and the third row represent the block position predicted and the robot joint angles generated by the RNNPB. Plot in the bottom row represents the parametric bias of the RNNPB.

Fig. 10. Information flows in the ball handling task (a) and in the block handling task.

Fig. 11. The transition of landscape of dynamical structure in the ball handling task (a) and in the block handling task (b).

Performance of behavior switch and the prediction error during it at each timing

Performance of behavior switch and the prediction error during it at each timing