IEEE/RSJ IROS 2023 Workshop “Learning Meets Model-based Methods for Manipulation and Grasping”

Building robots capable of dexterous interaction with objects to carry out fine manipulation tasks has always been a grand challenge in robotics. The non-smooth, brittle nature of manipulator-object mechanics, together with perceptual uncertainty, easily violate the assumptions of early planning and control methods. Furthermore, accurate physical modeling of complex or non-rigid mechanical systems requires large amounts of computations, which is incompatible with real-time control.

Such challenges led researchers to develop a wide range of approaches, from adaptive control tailored to the (potentially changing) properties of the object at hand, to advanced perception to tackle measurement uncertainty. Machine learning also contributed by providing actionable representations of complex geometries and visual appearance, and by encoding hard-to-model expert demonstrations to reduce the cost of trial-and-error. In turn, these informed the development of novel robot control methods enabling more robust and dexterous skills. At the same time, the employment of mechanical models proved effective for enforcing structural constraints in robot control systems (including learning-based ones), thus improving safety and guiding exploration.

However, there are still many open challenges that need to be addressed to achieve long-horizon robotic manipulation and sidestep the computational burden of accurate simulation of contact-rich scenarios. The ambition of this workshop is to provide a comprehensive overview of the broad and scattered state of the art in robot manipulation and grasping, spanning model-based and learning-based approaches. Talks and interactive sessions will enable a deeper understanding of current approaches in different use cases, while stimulating the development of new methods.