Imitation Learning with Additional Constraints on Motion Style using Parametric Bias

TL;DR

This work addresses the tendency of imitation learning to converge to a single average motion style by introducing parametric bias (p) within an RNNPB framework to encode and control motion style through soft constraints. By formulating a loss that combines state-prediction with constraints on muscle tension, muscle length velocity, and joint velocity, and by updating $\bm{p}$ either offline or online, the method enables deliberate shaping of motion style while preserving task reproduction. Experimental validation across a simulated tendon arm, PR2, and MusashiLarm demonstrates the ability to modulate velocity and force and to adapt to changing robot configurations, though challenges remain in balancing multiple constraints and extending to non-quantifiable styles. The approach offers a practical pathway to versatile, style-aware imitation in complex robotic systems, with potential for richer multimodal constraints and longer-horizon tasks in future work.

Abstract

Imitation learning is one of the methods for reproducing human demonstration adaptively in robots. So far, it has been found that generalization ability of the imitation learning enables the robots to perform tasks adaptably in untrained environments. However, motion styles such as motion trajectory and the amount of force applied depend largely on the dataset of human demonstration, and settle down to an average motion style. In this study, we propose a method that adds parametric bias to the conventional imitation learning network and can add constraints to the motion style. By experiments using PR2 and the musculoskeletal humanoid MusashiLarm, we show that it is possible to perform tasks by changing its motion style as intended with constraints on joint velocity, muscle length velocity, and muscle tension.

Figures (14)

• Figure 1: Imitation learning with additional constraints on motion style.
• Figure 2: The overall system of imitation learning with additional constraints.
• Figure 3: Experimental setup: simple movement by 1-DOF tendon arm, box opening by PR2, and box closing by MusashiLarm.
• Figure 4: Simple movement experiment by simulated 1-DOF tendon arm when changing $\bm{r}$ and $\bm{f}^{style}$: parametric biases trained in data collection phase and those updated from additional constraints regarding "A", "B-maximize", "B-minimize", "AB-maximize", and "AB-minimize".
• Figure 5: Evaluation of simple movement experiment by simulated 1-DOF tendon arm when changing $\bm{r}$ and $\bm{f}^{style}$: transition of $||\bm{\theta}^{task}-\bm{\theta}_{t}||_{2}$ and $||\bm{f}_{t}||_{2}$.