Considerable progress in grasping and manipulation has been achieved using approaches that extract complex behaviors from data. Yet, data-driven approaches are mostly assessed empirically and not necessarily complying with physical and dynamical constraints compared to analytical approaches where these constraints can be modeled manually. Besides, application of black-box learning models often results in limited success due to large data requirements, incompetence in yielding physically consistent results, and lack of generalizability to novel cases. Meanwhile, physics-based approaches have also been improved dealing with uncertainty. Yet, simplifying assumptions on, e.g. contact and friction model, stationary environment, are often needed resulting in models that cannot account for variations arising when contact models are rich or environments are unstructured and dynamically change. As neither a learning-based nor an analytic approach can be considered sufficient for complex manipulation tasks with high dimensional state spaces, a continuum between mechanistic and learning models is indispensable, where both domain-specific knowledge and data are integrated synergistically. In contrast to practices based on simple forms of feature engineering, heuristics, and constraints, this workshop is focused on exploring a deeper coupling of learning-based methods with physics and discussing benefits of analytical and data-driven approaches in grasping and manipulation applications.
Check out the workshop website.