Periodic Reporting for period 1 - RobotSL (Robot skill learning: imitation, exploitation and control)
Período documentado: 2021-04-01 hasta 2023-03-31
Imitation learning has emerged as an important research direction in the community of robot learning. Due to its nature and user-friendly features, imitation learning has achieved great success in a myriad of scenarios, e.g. surgical robots, agricultural robots, and human-robot collaboration.
It is well-known that robots often encounter various constraints, e.g. the robot’s end-effector must comply with the plane constraint when wiping a table or writing letters on a board. Moreover, robots must always meet their intrinsic constraints, e.g. motion range, joint and torque limits. In addition, environmental priors and multi-modal features will help to learn skills in complicated and unstructured scenarios.
The main goal of the project is to provide an imitation learning framework that can handle various external and internal constraints, and explore environment priors and multi-modal features.
The success of the project can bring direct benefits to society, such as equipping a robot with the friendly human-robot interaction ability to take care of the elderly and cope with daily tasks including cleaning floors and carrying heavy objects.
It is well-known that robots often encounter various constraints, e.g. the robot’s end-effector must comply with the plane constraint when wiping a table or writing letters on a board. Moreover, robots must always meet their intrinsic constraints, e.g. motion range, joint and torque limits. In addition, environmental priors and multi-modal features will help to learn skills in complicated and unstructured scenarios.
The main goal of the project is to provide an imitation learning framework that can handle various external and internal constraints, and explore environment priors and multi-modal features.
The success of the project can bring direct benefits to society, such as equipping a robot with the friendly human-robot interaction ability to take care of the elderly and cope with daily tasks including cleaning floors and carrying heavy objects.
In order to endow robots with a strong generalization ability, we first developed an algorithm capable of dealing with incomplete orientation constraints. For example, when a robot is trying to unscrew a bottle, its orientation along the vertical axis must comply with a strict restriction so as to unscrew the lid successfully, while the orientation along the other two axes is less important. By handling incomplete orientation constraints, our framework provides more flexibility for the robot when it tries to learn orientations from demonstrations.
Considering that the smoothness of a trajectory is crucial for a robot, we proposed an algorithm to generate smoother trajectories. Specifically, we exploited a linearly constrained nonparametric imitation learning method and extended it by incorporating the learning of acceleration curves. Since the higher-order trajectories are studied, the framework not only generates smoother trajectories but also permits constraints defined on acceleration.
While nonparametric imitation learning alleviates the explicit definition of basis functions, its computational cost is often expensive, which is cubed to the size of the training dataset. This limitation inevitably restricts the applications of the class of nonparametric approaches. Therefore, we leveraged the concept of null space into nonparametric learning, yielding a novel approach that dramatically reduces the computational time of nonparametric approaches.
Note that in real tasks nonlinear constraints are more useful than linear constraints, which can find various applications in broader settings, such as obstacle avoidance and self-collision avoidance. Therefore, a comprehensive framework was derived in this project, which was built on top of kernelized movement primitives. The framework can cope with both linear and nonlinear constraints, adaptation issues, as well as obstacle avoidance problems. Specifically, the framework still maintains the merits of nonparametric approaches.
Overview of results and dissemination
This project has led to two publications, one was published in IEEE/ASME Transactions on Mechatronics, and the other was published in the 2023 IEEE International Conference on Robotics and Automation (ICRA). In addition, one paper is under review and another three papers are in preparation that are expected to be submitted to journals.
The dissemination of the project has also been showcased in wide collaboration with other universities and institutes. For example, the ongoing joint work with the University of Southern Denmark to accomplish manipulation tasks in a surgical robot, where the transfer of expert skills to a surgical robot is used.
Considering that the smoothness of a trajectory is crucial for a robot, we proposed an algorithm to generate smoother trajectories. Specifically, we exploited a linearly constrained nonparametric imitation learning method and extended it by incorporating the learning of acceleration curves. Since the higher-order trajectories are studied, the framework not only generates smoother trajectories but also permits constraints defined on acceleration.
While nonparametric imitation learning alleviates the explicit definition of basis functions, its computational cost is often expensive, which is cubed to the size of the training dataset. This limitation inevitably restricts the applications of the class of nonparametric approaches. Therefore, we leveraged the concept of null space into nonparametric learning, yielding a novel approach that dramatically reduces the computational time of nonparametric approaches.
Note that in real tasks nonlinear constraints are more useful than linear constraints, which can find various applications in broader settings, such as obstacle avoidance and self-collision avoidance. Therefore, a comprehensive framework was derived in this project, which was built on top of kernelized movement primitives. The framework can cope with both linear and nonlinear constraints, adaptation issues, as well as obstacle avoidance problems. Specifically, the framework still maintains the merits of nonparametric approaches.
Overview of results and dissemination
This project has led to two publications, one was published in IEEE/ASME Transactions on Mechatronics, and the other was published in the 2023 IEEE International Conference on Robotics and Automation (ICRA). In addition, one paper is under review and another three papers are in preparation that are expected to be submitted to journals.
The dissemination of the project has also been showcased in wide collaboration with other universities and institutes. For example, the ongoing joint work with the University of Southern Denmark to accomplish manipulation tasks in a surgical robot, where the transfer of expert skills to a surgical robot is used.
The project focuses on endowing robots with versatile skills through learning a few demonstrations. This project covers imitation learning and data exploration, which will provide new research insights for the community of robot learning. Specifically, the project has the potential of reducing the gap between the transfer of expert skills and robot learning. Moreover, the applications of the project can be straightforward in many areas, such as rehabilitation training and human-robot collaboration.