🔥We enhance end-to-end robotic motion imitation systems by introducing a motion synthesis method, which optimizes the use of costly motion capture (mocap) data. This advancement addresses the challenge of controlling robots to perform specific tasks using only limited segmented motions and imitation learning.

🔥We propose a State Variational Autoencoder (SVAE) model that compresses the dynamic transformation between current and subsequent postures into a latent space. This model facilitates the generation of diverse and high-quality poses while avoiding consecutive repetitions.

🔥We develop a Control Network of Synthesized Motion (SMC-Net), trained through deep reinforcement learning (DRL), to seamlessly concatenate segmented motions and synthesize poses from various sources. This is achieved by simultaneously training multiple decoders within the SVAE framework, resulting in task-oriented behavior through cumulative rewards.

🔥We explicitly define Critical Joint Constraints (CJC) and incorporate a penalty based on these constraints into the original task-specific reward function. This modification enhances the naturalness and quality of movements. The effectiveness of our framework is demonstrated through its successful execution of various reach-target-and-reaction tasks.