Bionic Design is widely considered as a major source of inspiration in making advanced robots featuring merits from biological evolution for applied engineering.

  • We are particularly interested in a class of design through the interaction between rigid and soft materials for controllable, safe, and adaptive actuation by learning from the lobsters, shrimps, crabs, or crustaceans in general.
  • To this effort, we’ve invented a series of hybrid actuation mechanism where articulated rigid shells connected in various fashions are filled with different forms of soft chambers for collaborative actuation. Under fluidic actuation, the non-linear (hyper-)elastic deformation of the soft materials is bounded by the external rigid shells, which in turn, provides a robust and simple mechanism to provide kinematic and dynamic motions that can be analyzed using classical theories in mechanisms and robotics.
  • One of our major goal is to develop an untethered amphibian robot with almost one-to-one form factor as the actual lobster/crabs with comparable locomotion capabilities under water and on lands, a sophisticated dual-arm system for object manipulation, and a low-cost vision and sensing system for detection and navigation in blurring and muddy environment.

Robot Learning is an algorithm design that concerns the programming of robotic systems are capable of self-learning from interactions with the physical environment.

  • We are particularly interested in advanced robot grasping where arm-and-gripper type of robotic systems equipped with RGB-D vision systems to autonomously detect, recognize, plan, and execute object grasping and manipulation tasks.
  • A featured effort of ours is to develop a symbolic Arcade Claw Robot Grasping Framework, i.e. DeepClaw, that is shareable and reproducible as a physical benchmark to generate reusable learning data for robot learning disregard the differences in the robot hardware, especially when they are still relatively expensive to be mass adopted by the industry or for home use. While exploring algorithm designs based on human logical reasoning, we are also developing three versions of a DeepClaw platform including an entry version using LEGO parts, a standard version using industrial parts, and an advanced version using industrial-grade collaborate robot and vision hardware.
  • Through this DeepClaw platform, we intend to test and benchmark various robot learning algorithms developed by other labs, use it as a platform to generate specific dataset for robot learning that can be used across different platforms, and more importantly, explore the boundary of generalization in robot learning disregarding the limitation in robot hardware.

Copyright © 2018 All Rights Reserved.