NAME: Compliant Physical Interaction for Robot Manipulation Tasks

Description

Robotic manipulation of everyday objects and execution of household chores are among the most challenging skills for future service robots. Most of current research in robot grasping is limited to pick-and-place tasks, without paying attention to the whole range of different tasks needed in home environments, such as opening doors, interacting with furniture, electrical appliances, etc. The few robots exhibiting such abilities do it in an ad-hoc fashion with very precise models of the environment, and typically planning a trajectory for the end-effector based on such models. They are carefully programmed for the specific task and objects used in the experiments resulting in a very brittle execution. However, nowadays, complex dexterous hands are available as well as precise sensor devices.

We believe a methodology based on well-established robotics techniques and hardware going hand in hand with new systems design and engineering principles taking inspiration from findings in neuroscience will pave the way for complex manipulation tasks with very little knowledge about the environment, and without accurately planning any hand trajectory in advance. Task- and object-independent grasping algorithms and control laws must be identified, able to deal with very different objects and tasks under uncertainty, and without being specially designed for any particular kind of task, but relying on flexible control architetures based on multisensor feedback.

 

Objectives

The overall goal is to develop and implement the architecture and methods for a robot to be able to perform a variety of tasks involving compliant physical interaction in everyday domestic environments by means of a multifinger hand and using vision, force and tactile sensing. The system should be designed in such a way that tasks are executed robustly without precise models, and new tasks can be easily dealt with. Ideas for neuroscience findings will serve as inspiration.

1. To elaborate a theoretical framework for an integrated specification of both grasps and tasks, so that any grasp or physical contact is defined as a desired task-suitable relationship between the robot hand and the object being manipulated, whereas the task is defined under the task frame formalism as a desired motion that must be applied to the object. This framework should be amenable to describe a variety of tasks involving different kinematic mechanisms and allow sensor-guided compliant execution.
2. To design and implement a software/control architecture for the robot system, that is based on the previous framework and is suitable for easily defining and performing a variety of compliant manipulation tasks, as well as the incremental incorporation of more complex tasks in a modular fashion.
3. To develop techniques for sensor-guided compliant task execution, including sensor coupling — combining vision, force and tactile sensing— and a good estimation of the hand-to-object pose in execution time to allow the robot to adapt its motion to the particular kinematic constraints of the object, without the need of a detailed model.
4. To demonstrate the validity of the aproach by means of a series of controlled experiments with performance metrics and benchmarks of increased difficulty, in which the robot will have to deal with a variety of different domestic tasks involving compliant physical interaction such as: door/drawer opening under different conditions, taking books out of bookshelves, ironing or pushing objects on a table, inserting one object into another, gently setting down objects on work surfaces, reorienting objects for re-grasping, pushing buttons, flipping levers, turning taps, etc.