Learning Shared Encoding of Related Robotic Tasks

Conventional techniques usually solve the task transfer problem through bottom-up approaches or cleverly designed tricks, such as Reinforcement Learning or reverse execution. However, leveraging the common semantic context between executions of similar or related tasks may promote learning performance by enabling the utilization of shared representations. In this study, we propose a Learning-from-Demonstration model that simultaneously learns from semantically related demonstrations while actively attempting to discover the common representations of both. By doing so, our model learns to generate trajectories from related demonstrations. We showcased the performance of our model in synthetically generated trajectories. Later, we assessed our model with a daily, related task couple in a simulation environment.