Nicolas Gorges

Haptic object recognition using passive joints and haptic key features

This paper presents a novel approach for haptic object recognition with an anthropomorphic robot hand. Firstly, passive degrees of freedom are introduced to the tactile sensor system of the robot hand. This allows the planar tactile sensor patches to optimally adjust themselves to the objects surface and to acquire additional sensor information for shape reconstruction. Secondly, this paper presents an approach to classify an object directly from the haptic sensor data acquired by a palpation sequence with the robot hand - without building a 3d-model of the object. Therefore, a finite set of essential finger positions and tactile contact patterns are identified which can be used to describe a single palpation step. A palpation sequence can then be merged into a simple statistical description of the object and finally be classified. The proposed approach for haptic object recognition and the new tactile sensor system are evaluated with an anthropomorphic robot hand.

Opening a door with a humanoid robot using multi-sensory tactile feedback

This paper presents a multi-sensor based generic approach to opening doors for a dexterous robot. Once the handle has been located by a computer vision algorithm and properly grasped, we are able to open doors without using a model or other prior knowledge of the door geometry. This is done by combining the sensor information of both a force-torque sensor in the robot wrist and a tactile sensor matrix in the robot gripper itself. Our experimental results show that the combination of both sensors achieves the most successful way to open the door.

Tactile sensing for an anthropomorphic robotic hand: Hardware and signal processing

In this paper, a tactile sensing system for an anthropomorphic robot hand is presented. The tactile sensing system is designed as a construction kit making it very versatile. The sensor data preprocessing is embedded into the hands hardware structure and is fully integrated. The sensor system is able to gather tactile pressure profiles and to measure vibrations in the sensors cover. Additionally to the introduction of the hardware, the signal processing and the classification of the acquired sensor data will be explained in detail. These algorithms make the tactile sensing system capable to detect contact points, to classify contact patterns and to detect slip conditions during object manipulation and grasping.

Haptic object recognition for multi-fingered robot hands

In this paper, we present an approach for haptic object recognition and its evaluation on multi-fingered robot hands. The recognition approach is based on extracting key features of tactile and kinesthetic data from multiple palpations using a clustering algorithm. A multi-sensory object representation is built by fusion of tactile and kinesthetic features. We evaluated our approach on three robot hands and compared the recognition performance using object sets consisting of daily household objects. Experimental results using the five-fingered hand of the humanoid robot ARMAR, the three-fingered Schunk Dexterous Hand 2 and a parallel Gripper are performed. The results show that the proposed approach generalizes to different robot hands.

Haptic object recognition using statistical point cloud features

This work presents a point cloud approach for haptic object recognition with an anthropomorphic robot hand. It introduces several statistical point cloud features to provide robust descriptions of objects. It addresses the domain specific problems of sparsely populated and distorted point clouds that result from the direct interaction with the object. Also the contact normals registered during exploration — a natural byproduct — are taken into account for computing some of these features.

Haptic object exploration using attention cubes

This paper presents a new approach to generate a strategy for haptic object exploration. Each voxel of the exploration area is assigned to an attention value which depends on the surrounding structure, the distance to the hand, the distance to already visited points and the focus of the exploration. The voxel with the highest attention value is taken as the next point of interest. This exploration loop results in a point cloud which is classified using an Iterative-Closest-Point algorithm. The approach is evaluated in a simulation environment which includes in particular a simulation of tactile sensors.

A cognitive path-guidance-system for minimally invasive surgery

The presented path-guidance system is able to learn movements and to predict motion. It shall enhance safe navigation for surgeons in minimally invasive surgery by creating a virtual fixture which holds the end-effectors motion to a desired path and warning the surgeon in a dangerous situation. Surgeons can demonstrate interventions and best practices. The system collects information from surgeon demonstrated trajectories, defined as best practices, and extracts knowledge to provide guidance for other users to carry out the same intervention. Knowledge extraction is achieved through trajectory clustering, maximum likelihood classification and a Markov model to predict states. The fundamental task is to guide a surgeon along a desired trajectory (navigated path) and prevent them entering into zones of risk. The path is not sequential, furcations are permitted and modeled showing alternatives in the ongoing intervention. An evaluation with a pelvitrainer showed good results with over 89% hit rate in predicting the motion.

Reactive hand-arm coordination for a humanoid robot using extended petri-nets

This paper provides a new approach for the temporal coordination of a humanoid robot, in particular the coordination of hand and arm movements. Our approach uses Petri-nets extended by callback and trigger functions to describe the identified features of coordinated movements. The Petri-net is not only used for modelling but also for the implementation - it describes and executes coordinated movements. It keeps the distinct control systems simple and manageable while still providing a broad variety of possible tasks. It furthermore allows the exchange of sensor information between independent control systems and thus allows the system to react on sensor feedback.

A framework for creating, coordinating, and executing skills on a humanoid robot

This paper presents a new bottom-up framework for programming a humanoid robot which keeps the control systems simple and manageable while still providing a broad variety of possible tasks. Firstly, the framework implies an approach for the creation and execution of skills on a humanoid robot. A skill is a small, specialized, task specific object which contains not only data but also active components like algorithms which are able to interpret and modify the data under the consideration of the changing environment. Secondly, the framework provides an mechanism for the temporal coordination of a humanoid robot. This mechanism uses Petri-nets extended by callback and trigger functions to describe the identified features of coordinated movements. The Petri-net is not only used for modeling but also for the implementation - it describes and executes coordinated movements. The framework provides furthermore a mechanism for the resource management of a humanoid robot which allows the parallel execution of skills and the supervision of their execution. The proposed framework is evaluated in the sample application of a fetch and carry task requiring a coordinated movement of hand, arm and head.

Learning an Object-Grasp Relation for Silhouette-Based Grasp Planning

This work addresses the problem of experienced-based grasping of unknown objects. A relation between objects and grasps is learned based on example grasps for a set of given objects. This machine-learning approach is applied to the problem of visual determination of grasp points based on a given silhouette of an object. The approximated function allows computing a grasp quality for any objectgrasp combination. For the dimension reduction of the object silhouettes, a combination of image normalization and principal component analysis is used. A Support Vector Regression is used to learn the object-grasp relation. For the evaluation, the objects are grasped with a two-finger gripper and an imprint of a planar object on a tactile sensor matrix is used as an imaging method.