My major research direction is learning control for fine manipulation tasks. Manipulation involves robot hands that are exceedingly complex due to their dynamics and under actuation and further the challenges associated with prehensile and non-prehensile interactions. To reduce the controller complexity of dexterous hands, the idea of postural synergies is widely adopted but their motivation shadows down with the increasing complexity of manipulation tasks. In addition to this, enabling robots to acquire human-level capabilities necessitates to exploit multi-sensory information for autonomous contact-rich manipulations.
To alleviate such problems, a new framework called kernelized synergies is developed, which is capable of exploiting the same reduced subspace for complex robust grasping and dexterous manipulation tasks. Moreover, for autonomous interaction with distinct objects, visual and tactile perceptions are integrated into proposed framework for object localization and run-time force adaptation respectively. In addition to this, kernelized synergies framework is also used to initialize the control policies of different RL algorithms (DAPG, PI2, POWER etc) to speed up the learning process on exploring the state-action space for collaborative human-robot manipulations.
1. Best Research Paper Award by Franka-Emika @ HFR2020