Dexterity in Robotic Grasping, Manipulation and Assembly
- Author(s): Fan, Yongxiang;
- Advisor(s): Tomizuka, Masayoshi;
- et al.
Industrial manipulators are programmed and integrated into different systems to deliver various functions. Traditional industrial manipulators are highly efficient and precise in mass production but deficient in flexibility and dexterity in mass customization due to the heavy reprogramming efforts in limited product life cycle, variations of environments and uncertainties during robot-environment interactions.
This dissertation aims to address the aforementioned deficiencies by improving the dexterity of industrial manipulators. The manipulators with proposed algorithms are required to 1) reduce the hand-engineering in end-effector design, parameter tuning and system integration, and 2) exhibit robustness to uncertainties during the interaction with environments. The fulfillment of the requirements is decomposed into three aspects in this dissertation. The first aspect is to realize kinematic dexterity by developing a unified grasping framework with both customized end-effectors and general hands on objects of different categories. The second aspect is to achieve dynamic dexterity by constructing an in-hand manipulation and finger gaiting architecture to manipulate the grasped objects robustly and precisely. The third aspect is to attain skill dexterity by designing an intelligent assembly algorithm to learn assembly skills from uncertain environments.
The developed grasping framework actively avoids collision and is able to plan grasps and trajectories efficiently with different hands. The grasp planning with industrial customized grippers by surface fitting is introduced in Chapter 2, and the planning efficiency is improved by a learning-based grasp explorer in Chapter 3. The transferring of grasps from parallel grippers to multi-fingered hands by finger splitting is discussed in Chapter 4. Chapter 5 further presents an optimization model to directly plan precision grasps with multi-fingered hands. The final grasping framework is proposed in Chapter 6 by combining the optimization model with a multi-dimensional iterative surface fitting, to improve the grasp versatility and robustness of the optimization model. The constructed manipulation architecture achieves robust grasping and dexterous manipulation under uncertainties. A comprehensive architecture is introduced and verified with different physical multi-fingered hands in Chapter 7. Chapter 8 proposes a robust manipulation controller within the architecture to further increase the robustness under various uncertainties. To relocate fingers for long-range object motion, the manipulation architecture is augmented with a high-level finger gaits planner in Chapter 9. The designed assembly scheme learns automatic assembly skills with the proposed guided-deep deterministic policy gradient (guided-DDPG) in Chapter 10. By combining supervised learning and reinforcement learning, the proposed guided-DDPG is more efficient than reinforcement learning and achieves better stability and robustness compared with supervised learning.
The proposed grasping and manipulation strategies with customized/general-purposed grippers are able to reduce hand-engineering in mass customization and extend dexterities of automation systems in both kinematic and dynamic levels. The proposed learning assembly scheme increases the efficiency and stability of the assembly in contact-rich scenarios and extends the dexterity of automation systems in skill level. The effectiveness of the grasping, manipulation and assembly algorithms are verified by a series of simulations and experiments on different manipulators and hands.