The results obtained by Fanny Ficuciello during her research activity, published in journals and proceedings of the robotic international community, are here resumed.
Control of a complex hand-arm robotic system during grasping
The problem of controlling a complex hand-arm robotic system during grasping tasks has been addressed in [IC-5]. Such a system can interact with the environment or with a human being, for example during the exchange of an object. The applied control is in charge of ensuring that the hand firmly grasps the object and that the arm is compliant with respect to external forces applied to the grasped object. Relying on grasping theory, the reconstruction of the external forces applied to the object by means of the force measurements at the fingertips is used in the control action. Force measurements are also used to ensure the adjustment of the internal forces exerted on the object to have a stable grasp during the interaction with the environment.
Control of two arms robotic system during grasping
With regards to safe human-robot interaction control, an algorithm, that enhances safety and reliability, based on the impedance control of a two arms system that grasps an object and interacts with the environment, has been developed in [IC-6]. To obtain a safe behavior of the system during the intentional interaction with the object and unintentional interaction on the body, a compliant behavior of the entire system is ensured by applying an impedance control at three levels, i.e. object, end-effector and body level. A centralized impedance control strategy is provided to confer a compliant behavior to the object, while an active decentralized impedance with force tracking control is enforced to the end-effectors to control internal forces on the object. Furthermore, a compliant behavior on the body of the dual-arm system is obtained by means of null-space impedance control while the manipulability measure of the whole system is locally maximized.
Control of anthropomorphic hands
Fanny Ficuciello has developed techniques of planning and control based on the notion of postural synergies and neural networks inspired by neuroscience studies carried out on human hand and the functioning of the human brain. In [IJ-1, IC-2, IC-3, IC-4], different techniques to obtain the synergies subspace of an anthropomorphic robotic hand, using the human hand as a guide, have been tested and compared. In [IJ-2, IC-7, IC-8], a set of grasping postures performed by five subjects in grasping commonly used objects has been mapped to a robotic hand assuming its own kinematics as a simplified model of the human hand. Using an RGB camera and depth sensor for 3D motion capture, the human hand palm pose and fingertip positions have been measured for the reference set of grasps. In [BC-3] a neural network model has
been designed for planning grasps of a cybernetic hand prototype by means of postural synergies. In [IC-9], the problem of in-hand dexterous manipulation has been addressed on the base of postural synergies analysis.
Port-Hamiltonian Modeling for Soft Manipulation
In [IC-1], the theory of port-Hamiltonian systems has been explored to model a multifingered robotic hand, with soft-pads, while grasping and manipulating an object. The theory of port-Hamiltonian systems allows to describe the system behavior in a coordinate-free way and can be naturally extended to include constrained systems and compliant contact models. The viscoelastic behavior of the contact is described in terms of energy storage and dissipation. Using the concept of power ports, the dynamics of the hand, the contact, and the object are described in a coordinate-free way. Moreover, an IPC is applied to control the motion of the object and to regulate the internal forces. The main advantage of the port-Hamiltonian formulation for constrained systems is that we do not need to modify the dynamic equations when a change occurs in the contact state. Instead, it is possible to represent both cases in a time-dependent geometrical structure, that satisfies the power continuity conditions in every contact state. This framework allows to approach the problem in a more intuitive and compact way.
Learning and Control Strategies for Underactuated Anthropomorphic Hands
Underactuated hands require the investigation of planning and control methods that disregard accurate definition of the desired contact points on the object and guarantee robustness with respect to variability of shape and size. To overcome these problems, Fanny Ficuciello has developed control and learning algorithms that rely on the synergies concept. The contribution is divided in three parts concerning synergies
computation, control and learning. Those parts are different aspects of the same research that goes towards the realization of human-like prehensile capabilities and autonomous learning skills for a robotic upper limb system with anthropomorphic design. In [IC-13], the same data set of grasps, measured on five human subjects and available from [IC-8], is used to evaluate the grasping capabilities of an underactuated robot hand in a synergy-based framework. Once the synergy matrix of an underactuated hand has been computed, in order to test the efficiency of the mapping method, different grasps can be reproduced in the three dimensions synergies subspace. Actually, since mechanical synergies of the underactuated hand affect the mapping from the human hand, the projection of a grasp from the data set in the synergies subspace is not so effective as for the full-actuated anthropomorphic hands. This means that a control strategy is required to adjust the reference grasp in order to let the hand adapting to the object while moving in the synergies subspace. The synergies subspace has been tested for the hand control using a kinematic algorithm based on has been adopted. The fingertips desired positions are modified in the control algorithm in order to reduce their distance with respect to the centroid of a virtual object computed as the centroid of the fingertips involved in the desired grasp. Moreover, in order to limit the grasping forces, the desired target is further modified on the basis of the measured motor current and of a defined threshold that is related to the texture of the object. The experiments demonstrate that the synergies subspace is suitable for hand control in grasping a wide variety of objects, i.e. the algorithm is stable and effectively regulates the grasping forces by modifying the motor positions in the synergies subspace. Moreover, to improve the grasping capabilities strategies based on quality indexes to close the hand toward the object in the synergies subspace have been developed in [R-2]. Model-based control strategies relying on synergistic models of manipulation activities learned from human experience can be integrated with real-time learning from actions strategies. The use of supervised learning, such as artificial neural networks, or reinforcement learning techniques, serves for the parameterization of synergies depending on task requirements. Fanny Ficuciello has developed Supervised Learning (SL) and Reinforcement Learning (RL) stragegies. In [C-18] a supervised learning strategy has been applied in conjunction with a control strategy to provide anthropomorphic hand-arm systems with autonomous grasping capabilities. Both learning and control algorithms have been developed in a synergy-based framework in order to address issues related to high dimension of the configuration space, that typically characterizes robotic hands and arms with human-like kinematics. An experimental setup has been built to learn hand-arm motion from humans during reaching and grasping tasks. Then, a Neural Network (NN) has been realized to generalize the grasps learned by imitation. Since the NN approximates the relationship between the object characteristics and the grasp configuration of the hand-arm system, a synergy-based control strategy has been applied to overcome planning errors. In [C-17] is demonstrated that a synergy based approach is powerful for reinforcement learning of grasping with anthropomorphic hands due to configuration space dimesionality reduction that guarantees the convergence and efficiency of the learning algorithm. The design of appropriate policy representations is essential for RL methods to be successfully applied to real-world robots. PCA and human grasps data set serve as data structures to define a policy and its initial parameters for a RL algorithm. Indeed, starting from a good enough demonstration, the algorithm can optimize the policy parameters to gradually refine a stable grasp. When a clear measure about the success of the task is available, RL adaptability to new objects is ensured. A key point is the adoption of a suitable reward function representing the goal of the task and ensuring one-step performance evaluation.
Impedance control strategies with redundancy resolution for human-robot intuitive co-manipulation
The control of safe and efficient physical interaction between a human and the robot co-worker has been addressed in [IC-10, BC-4] by investigating new and efficient approach to redundancy resolution and variable impedance control strategies. Redundancy has been exploited to make the robot equivalent inertia at the end-effector as close as possible to the desired inertia. In particular, since co-manipulation tasks typically require a decoupled impedance along the Cartesian directions, the redundant degrees of freedom are used to reduce as much as possible the dynamic coupling of the end-effector equivalent inertia. In [IJ-3, IJ-5, IC-11], a study on variable impedance control for a redundant robot arm to enable an intuitive and safe physical interaction with humans during the execution of co-manipulation tasks has been proposed. The impedance parameters are modulated on-line according to the human behaviour during the interaction.
Surgery simulators with realistic haptic rendering
Recently, the research activities of Fanny Ficuciello have been extended also to surgical robotics. In [IC-14] a soft-rigid collision algorithm has been integrated in an open source physics engine, Bullet Physics. In surgical applications this can be the case of a clamp grabbing deformable organic materials or of a spatula opening a brain fissure. The default soft-rigid collision algorithm proposed in Bullet is not very effective in the
case of thin tools interacting with deformable objects. In particular, if the rigid body (surgical tool) moves slowly, i.e. its displacement covers a small distance compared to the simulation step size, the collision is detected regularly, otherwise the default algorithm does not recognize the collision. As a consequence, the object penetrates into the soft body. Besides the implementation of the soft-rigid collision algorithm, the new contribution consists on generalizing the algorithm to different shaped rigid object such as convex rigid bodies with thin thickness along one of the three main directions. Moreover, the haptic rendering has been realized by controlling the spatula in the 3D virtual space with the Novint Falcon 3D Haptic Controller. The default linear elastic model of the interaction force has been replaced with a more realistic and physical consistent non-linear viscoelastic model. As a second step, the algorithm has been further extended to a clamp constituted by two rigid colliding objects grabbing deformable materials.