Keywords: Co-design, hardware optimization, manipulation, in-hand
TL;DR: Co-design for in hand rotation via Cross Entropy MPC and Baysian Optimization of hardware parameters
Abstract: In-hand manipulation for robots has recently been
possible due to advancements in reinforcement learning and
ongoing development of new robotic hands. Both advancements
have iteratively pushed the frontiers of manipulation with new
controllers allowing more complex manipulators to be effective
and hardware advancements allowing true dexterous manip-
ulation. Co-design aims to build co-optimized hardware and
control systems for complex tasks. This work aims to explore co-
design by optimizing hardware of robotic hands. Reinforcement
learning has commonly been used as a controller but is often
computationally expensive and time consuming to train - making
scaling resource intensive across various designs. To explore more
effective manipulators at scale, we propose a framework for joint
hardware optimization with Cross Entropy MPC for improved
sampling efficiency and hardware optimized manipulation.
Submission Number: 6
Loading