Cooperative Manipulation Strategies for Multi-Robot Collaboration

The demand for robots that are capable of performing complex tasks has soared in recent years; research interests in multi-robot systems have also risen. Despite significant advances in single-arm manipulation, the need for cooperative manipulation cannot be overlooked. This thesis presents strategies that enable multiple robots to perform collaborative tasks. More specifically, this thesis will address cooperative manipulation under the possibility of external disturbance. I propose two approaches: an impedance-based approach and a residual-reinforcement-learning-based approach. With respect to the impedance-based approach, I will focus on the concept of applied internal stress, on the jointly manipulated object, and how internal stress can be leveraged as a compensation mechanism for disturbance on a cooperative manipulation setup. I will present an impedance-based strategy that is used to determine how much compensation should be applied to the system. Additionally, I will present a decentralized approach to cooperative manipulation that is based on a residual reinforcement learning scheme. Residual reinforcement learning involves the superposition of a learned policy with a standard controller, in such a way that the input to the learned policy is informed by the output of the standard controller, or vice versa. I will show how a robot can compensate for unexpected partner behaviors without communicating with its partner(s). Finally, experimental demonstrations both in simulation and on the physical system will be presented. Although this work describes a multi-robot scenario, the experimental validation will be on a two-robot system. The physical system also includes mobile platforms that extend the workspace of the manipulators shown in simulation. The contribution of this thesis can be summarized as follows: 1. A collision reaction strategy based on internal stress loading 2. An impedance-based collision reaction strategy 3. Hardware demonstration of the above contributions 4. A cooperative manipulation approach via residual reinforcement learning

Date

2022-08-26

Resource Type

Text

Resource Subtype

Dissertation

Full item page