Self-Supervised Bayesian Visual Imitation Learning Applied to Robotic Pouring

Humans can learn to perform new tasks by watching a few examples of experts' demonstration videos and applying the learned knowledge to different scenarios. To enable robots to have a similar ability, we propose a self-supervised Bayesian visual imitation learning approach, which enables the robot to learn manipulation skills by watching humans from a third-person view. We chose the robotic pouring task as an example to evaluate the effectiveness of the proposed method. First, a spatial-temporal feature embedding (STFE) model is trained in a self-supervised learning manner, which encodes the prior knowledge obtained from human demonstration. Following that, the STFE model is fine-tuned using a few groups of demonstration data collected via teleoperating the robot for feature extraction, while an action generation model is trained for task execution with the embedded features of visual observation data as input. To further enhance the success rate of task execution, an adaptive safety boundary is constructed based on Bayesian inference to regulate the robot's action during the online deployment. Experimental results indicate that with the proposed method, automatic pouring can be achieved with a high success rate while data efficiency can be ensured
^1
1
For more details, please view our project website: https://sites.1google.com/view/rsspouring/home/.