Among the ways we teach robots to move, one involves learning from human examples so-called learning from demonstration (LfD). The idea is to mimic and understand complex movements by observing humans. This involves gathering information about various aspects of a motion trajectory such as positions, orientations, velocities, and so on. However, this blend of information makes it difficult for traditional methods, which assume simple structures such as Euclidean, to handle the complex way data is represented, and the high dimensionality of the involved data. To tackle this, we have combined the robot motion learning with advanced geometric methods.
This talk will cover, first how we use variational auto-encoders to learn a Riemannian manifold in a low dimensional latent space that represents the underlying geometry of the motion trajectories.
Furthermore, we will showcase how complex robot motion skills can be retrieved from the geodesics on these learned manifolds. Moving on, we shift our focus to the significance of learning dynamics. Instead of concentrating solely on position and orientation trajectories, we emphasize the importance of learning velocity trajectories. This significantly enhances the robot's capability to effectively cope with external disturbances, ensuring smoother adaptation to unexpected events.
Additionally, we will discuss how these learned dynamic systems achieve stability by conforming to contraction constraints on their Jacobian. Simultaneously, we will highlight the integration of a modulation matrix, which aids the robot in evading obstacles to ensure safety.
Hadi Beik Mohammadi is Doctoral Student at Bosch Center for Artificial Intelligence. Learn more here