The present invention relates to a robot system and to a method of operation therefore. Robots designed to directly collaborate with humans in a PFL (power and force limited) manner enable safe physical human-robot interaction. In such cas es, contacts between humans and the robot are like ly. These contacts can be of different kinds, i.e. they can be accidental or they can be intended by the human. In either case, in a robot arm in which several links are interconnected by joints, the contact will cause the angular velocity of at least one joint and/or the torque to which the joint is subject to deviate from an expected value, which may have been measured in a previous iteration of a programmed movement of the robot arm, or which may be calculated based on known weights and leverages of the robot links and a current posture of the ro- bot arm.

Conventionally, any significant deviation from the expected value will cause the robot arm to stop. In that way, if the deviation is due to a collision with the person, injury to the person can be avoid ed. In a co-operative working environment, a person may deliberately touch the robot, e.g. intending to guide it around a new obstacle that isn't taken ac count of in a motion program the robot is currently executing. In that case, stopping is not an appro priate reaction to the contact. It would therefore be desirable if a robot system was capable of distinguishing between accidental and deliberate contact.

An approach to this problem might rely on low-pass and high-pass filtering of joint load torques, based on the assumption that collisions contain more high frequency content and intended interac- tions contain more low frequency content. However, the classification quality is often not sufficient as other quantities (e.g. joint speed) and features (e.g. shape of rising edges) have to be taken into account for a successful classification. Since many of these quantities are dependent on physical pa rameters of a given robot, rules on how to take these parameters into account have to be developed individually for different robot models, requiring considerable investment in highly skilled labor.

The object of the present invention is therefore to provide a robot system and a method of operation therefore which provide a simple and economic way of distinguishing between accidental and deliberate contact.

According to one aspect of the invention, the ob ject is achieved by a robot system comprising a robot arm comprising a plurality of links and of joints by which one of said links is connected to another, means for determining, for any one of said joints, its angular velocity and the torque to which the joint is subject,

a neural network which is connected to said means for receiving therefrom angular velocity and torque data, and which is trained to distinguish, based on said angular velocity and torque data, between a normal operation condition of the robot arm, a con dition in which the robot arm collides with an out side object and a condition where a person deliber ately interacts with the robot arm.

In a preferred embodiment, the neural network is trained to determine the point of contact where the robot arm collides with an outside object and/or where a person deliberately interacts with the ro bot arm.

In a preferred embodiment, the neural network is trained to distinguish, based on said angular ve locity and torque data, between a normal operation condition of the robot arm, a condition in which the robot arm collides with an outside object and a condition where a person deliberately interacts with the robot arm and to determine the point of contact where the robot arm collides with an out side object and/or where a person deliberately in teracts with the robot arm. Such a point of contact could for example be localised below or above the elbow .

The robot system according to the invention is con figured to apply contact classification, and/or contact localization and/or combined classification and localization of contacts.

The robot system according to the invention easily scales to increasing the number of classification outputs. So for instance instead of "normal opera- tion"/"interaction"/"collision" the robot system can be configured to apply a classification such as "normal operation"/"upper arm interaction"/"upper arm collision"/"lower arm interaction"/"lower arm collision" . If the neural network is trained on a sufficient number of example data representative of the vari ous conditions to be distinguished, specific traits of the various conditions translate into weights of interconnections between the neurons of the net work, without the need for a human to actually rec ognize these traits and to formulate rules. There fore, a reliable distinction between collision and deliberate contact can be implemented at low cost in diverse robot systems regardless of their physi cal characteristics.

In order to facilitate a comparison between data sets and recognition of their common or different traits, data which the neural network receives as input in each time step has to have a constant num ber of elements. However, while a movement of the robot arm proceeds, the amount of data increases continuously. Therefore, the neural network is preferably designed to operate on a time series of angular velocity and torque data pairs of a given joint, wherein, in order to maintain the data set at a constant size, when a new data pair is deter- mined and added to the time series, the oldest pair is deleted.

The data thus obtained may be regarded as similar to video image frames in which e.g. data obtained at a given instant in time at the various joints form a row of pixel data, and a column is formed by successively obtained data of a given joint. In such a data set, recognition of the traits indica tive of collision or of deliberate contact is com- parable to pattern recognition. Theoretically, best distinction quality should be achievable if all neurons - or at least a large number of the neurons - of the neural network re ceive data from each joint. Therefore, in a basic embodiment of the invention, at least one neuron of the neural network is connected so as to receive angular velocity and torque data of each one of said joints.

However, the more connexions the neurons have, the longer it takes to train the network, or rather, the larger must be the quantity of training data. Therefore, it may be preferable not to allow all neurons to receive data from all over the robot arm, but to associate neurons to a specific joint of the robot. More precisely speaking, at least one neuron of the neural network may be connected so as to receive angular velocity and torque data of an associated one of said joints only.

In practice, each joint of the robot arm may have one or multiple neuron associated to it which re ceives data only from this one joint. Specifically, the neural network may comprise a first hidden layer which is divided into a plurali ty of groups, neurons of a given group being con nected so as to receive angular velocity and torque data of an associated one of said joints only, and being unconnected to neurons of other groups.

In that case, a second hidden layer may be provided whose neurons are connected to the various groups in the first hidden layer, so that while the neuron groups of the first hidden layer may recognize traits in the data from a given joint that might be indicative of accidental or deliberate contact, the neurons of the second hidden layer can provide an overview of the entire robot arm.

In a neural network in which each group is associ ated to one of said joints, each group can be trained individually to distinguish between normal operation of the associated joint, a condition in which the joint collides with an outside object and a condition where a person deliberately interacts with the joint. Since each group of neurons can fo cus on the data from its associated joint, recogni tion of traits indicative of a given type of con tact tends to be easier for such a group than for a network as defined above, in which neurons can re ceive data from any joint, so that good training results can be achieved using only a moderate amount of training data.

The robot system may further comprise a controller for controlling motion of the robot arm, the mode of operation of which is variable depending on an operation condition signal output by the neural network .

One of the operation modes of the controller may e.g. be a leadthrough mode, in which the control ler, based on the torque and angular velocity data from the joints, detects the direction of a contact force applied to the robot by a person, and con trols the robot to move in the direction of this force. Such a switchover will enable the person to e.g. guide the robot arm around an obstacle in its environment of which the controller is not or not yet aware, and thus to prevent a collision with the obstacle . Another one of the operation modes of the control ler may be a normal mode in which the movement of the robot arm is defined by a predetermined pro gram. The controller may enter said normal mode if the neural network finds that deliberate interac tion has ended, thus enabling the robot arm to im mediately resume normal operation as soon as the person stops guiding it. The robot system according to the invention can al so be configured to classify other types of contact situations. An example for such a contact situation would be a robot performing an assembly task, con tinuously measuring the position of its tool and the contact forces and moments that occur. From the course of positions as well as forces and moments can then be classified in the robot system accord ing to the invention with the neural network, whether the assembly task was completed successful- ly. The training of the neural network in the robot system for such a use case is equivalent to the training of the neural network for distinction and / or localization of human-robot contacts, just the input data and the outputs of the network change; in the case described so far: joint moments and ve locities in, classification of the contact out, in the second described case: Cartesian position and forces / moments in, classification of the success of the assembly task out.

According to a second aspect of the invention, the above object is achieved by a method of operating a robot system, the method comprising the steps of a) providing a neural network having a plurality of inputs, each of which is associated to a joint of a robot arm of said system, for receiving torque and angular velocity data of said joint, b) training said neural network to distinguish, based on said angular velocity and torque data, be tween a normal operation condition of the robot arm, a condition in which the robot arm collides with an outside object and a condition where a per son deliberately interacts with the robot arm based on time series of torque and angular velocity data representative of normal operation, of at least one collision and of a person deliberately interacting with the robot arm;

c) inputting real-time torque and angular veloci ty data of the robot arm into the neural network in order to determine therefrom an operation condition of the robot arm, and choosing an operation mode of the robot arm based on the operation condition.

In a preferred embodiment, the method further com prises the step of training the neural network to determine the point of contact where the robot arm collides with an outside object and/or where a per son deliberately interacts with the robot arm.

In a preferred embodiment, the method further com prises the step of training the neural network to distinguish, based on said angular velocity and torque data, between a normal operation condition of the robot arm, a condition in which the robot arm collides with an outside object and a condition where a person deliberately interacts with the ro bot arm and to determine the point of contact where the robot arm collides with an outside object and/or where a person deliberately interacts with the robot arm. Such a point of contact could for example be localised below or above the elbow.

The method according to the invention is configured to provide for contact classification, and/or con- tact localization and/or combined classification and localization of contacts.

The method according to the invention easily scales to increasing the number of classification outputs. So for instance instead of "normal opera- tion"/"interaction"/"collision" the robot system can be configured to apply a classification such as "normal operation"/"upper arm interaction"/"upper arm collision"/"lower arm interaction"/"lower arm collision" .

The method according to the invention can also be configured to classify other types of contact situ ations. An example for such a contact situation would be a robot performing an assembly task, con tinuously measuring the position of its tool and the contact forces and moments that occur. From the course of postions as well as forces and moments can then be classified using the method according to the invention with a neural network, whether the assembly task was completed successfully. The meth od is equivalent to the distinction and / or local ization of human-robot contacts, just the input da ta and the outputs of the network change; in the case described so far: joint moments and velocities in, classification of the contact out, in the sec ond described case: Cartesian position and forces / moments in, classification of the success of the assembly task out.

Further features and advantages of the invention will become apparent from the subsequent descrip tion of embodiments thereof referring to the ap pended drawings . Fig. 1 is a block diagram of a robot system ac cording to the present invention;

Fig . 2 illustrates a data which the neural net work of the robot system uses as input;

Fig . 3 is a first embodiment of the neural net work; Fig. 4 is a second embodiment of the neural net work; and

Fig. 5 is a third embodiment of the neural network. Fig. 1 illustrates a robot system comprising a ro bot arm 1 and its associated controller 2. The ro bot arm 1 comprises a support 3, an end effector 4 an arbitrary number of links 5 and joints 6 which connect to the links 5 to each other, to the sup- port 3 or to the end effector 4 and have one or two degrees of rotational freedom. As usual in the art, motors for driving rotation of the links 5 and the end effector 4 about axes 7, 8 are hidden inside the links 5, the joints 6 or the support 3. The joints 6 further comprise rotary encoders or other appropriate sensors 9 associated to each axis 7, 8 which provide the controller 2 with data on the orientation and angular velocity of each link 5, and torque sensors 10 which are sensitive to torque in the direction of axis 7 and 8, respectively. When the robot arm 1 is moving freely, without con tact to outside objects, the torque detected by these sensors 10 is governed by the weight and ge ometry of the links 5, their internal friction and, when the angular velocity is not constant, by their moment of inertia, so that the controller 2, based on known angles and rotational velocities of the links 5, can calculate an expected torque at each j oint .

If the torque detected by the sensors 10 deviates significantly from such an expected value, it can be assumed that the robot arm 1 is in contact with some outside object or person. Output data from the sensors 9, 10 are received by a neural network 11, which, in turn, determines an operating mode of controller 2.

A FIFO storage 12 is connected between the outputs of the sensors 9, 10 and the neural network 11, making available to the neural network 11 not only current torque and angular velocity data, but the M most recent sets of data from the sensors 9, 10, wherein M is an arbitrary integer.

The controller 2 has at least three operating modes, namely a normal operating mode in which it controls the robot arm to move according to a pre defined program, e.g. so as to seize a screw 13 and to introduce it into a workpiece 14. Another is an emergency stop mode in which the robot arm 1 is im- mediately brought to a halt, and a third one is a leadthrough mode in which the robot arm 1 will move into a direction into which it is urged by an out side force, e.g. by a user's hand 15. Fig. 2 is an "image frame" formed by angular veloc ity and torque data provided by the sensors of ro bot arm 1. The image frame comprises "pixels" 16, each of which corresponds to one data from one sen sor 9 or 10. These pixels 16 are organized in 2N lines, N being the number of degrees of freedom of the robot arm 1 and of angular velocity and torque sensors 9, 10 associated to each of these degrees of freedom, and M columns, each column correspond ing to a set of data obtained from said 2N sensors 9, 10 at a given instant in time and stored in the FIFO storage 13.

The task of the neural network 11 is to recognize, in an image frame formed by data from the sensors 9, 10, traits which are characteristic of acci dental and deliberate contact, so that when it rec ognizes accidental contact, it will switch the con troller 2 into emergency stop mode, and that, when it recognizes deliberate contact, the controller 2 is switched to leadthrough mode.

The neural network 11 is trained off-line to recog nize these traits by inputting training data ob tained from the robot arm in normal operation, in case of accidental contact and in case of deliber ate contact, and by optimizing coupling coeffi cients between the neurons of the network in a way known per se, by backpropagation, until the relia bility with which these different conditions are recognized by the network 11 is satisfactory.

The neural network 11 can have the structure shown in Fig. 3: In a hidden layer 17, there are P neu rons 18, each of which receives torque and angular velocity data from the M most recent data sets stored in storage 13, i.e. which is capable of "seeing" every pixel in the frame of Fig. 2. An output layer 19 comprises at least three neurons 20, one for each of the operating modes of control ler 2 which the network 11 has to choose from. Each of the neurons 20 receives input from all neurons 18 of hidden layer 17. The neurons 18 in the hidden layer 17 work with a standard sigmoid activation function while the neurons 20 in the output layer 19 use a softmax activation function. Only one of the neurons 20 can be active at a given time. When such a neuron 20 goes active, it switches control ler 2 into its associated operating mode.

Since each neuron 18 receives N*M input data, its vector of weighting coefficients must have N*M com ponents, so that the training process involves op timizing P*N*M weighting coefficients. In order to prevent the neural network 11 from simply memoriz ing its training data and the desired recognition results associated to them, a huge amount of train ing data is required, and the amount of computation required for satisfactory training increases far faster than in linear proportion with the number N of degrees of freedom of the robot arm 1.

According to a preferred embodiment shown in Fig. 4, the neural network 11 has a structure that re flects the structure of the robot arm 1: the neu rons 18 of the hidden layer 17 are divided into N groups 21, each of which is associated to one de gree of freedom of the robot arm 1, or to its cor responding joint 5. Neurons of one group 21 receive data only from the angular velocity sensor 9 and the torque sensor 10 of said one joint 5. This re duces the number of weighting coefficients in each neuron 18 by a factor of N. Since each group 21 su pervises only one joint 5, the number of neurons needed in one group 21 will be much less than the P neurons of Fig. 3; rather, since the complexity of the task the neural network 11 is to solve is the same in both cases, the total number of hidden lay er neurons 18 can be the same in both embodiments. So the amount of computation required for training is reduced not only due to the smaller number of weighting vector components, but also because the amount of training data needed to prevent the net work 11 from memorizing is smaller than in the case of Fig. 3. The structure of the network 11 shown in Fig. 5 is modelled on the fact that the effect of a contact, deliberate or accidental, on a given joint 5 de pends on where in the robot arm 1 the contact oc curred. Any contact will usually have the most no- ticeable effect on the joints that are immediately adjacent to the link where the contact occurred. Therefore, in a further preferred embodiment, training data are labelled not only as correspond ing to normal operation, accidental contact or de- liberate contact, but in the latter two cases, they also specify the link in which the contact oc curred. Obviously, for neurons 18 associated to a given joint in the hidden layer 17, it is much eas ier to recognize a condition in which a contact oc- curs with a link immediately adjacent to that joint than one where contact occurs with a faraway link. Therefore, in the embodiment of Fig. 5, the network 11 is divided into N sub-networks 22, each of which is associated to one joint or degree of freedom of the robot arm 1. Each sub-network 22 comprises a hidden layer 23 whose neurons 18 receive data from the angular velocity sensor 9 and torque sensor 10 of said joint only, just like those of the groups

21 of Fig. 4. In contrast to these groups 21, how- ever, the sub-networks 22 do not have a common out put layer, but each sub-network has an output layer 24 of its own.

Based on training data which indicate a link in which contact occurred, these output layers 24 can be trained to distinguish between normal operation and a condition where a contact, deliberate or ac- cidental, occurred in link adjacent to their asso ciated joint. Since a contact will usually produce the most clearly noticeable effect in a nearby joint, learning its characteristic traits is easier for the sub-network 22 associated to that joint than if faraway contacts had to be taken account of, too. Therefore, a good quality of distinction can be achieved here based on a small amount of training data.

The number Q of neurons in the output layers of the sub-networks 22 can be equal to the number of oper ating modes of controller 2. Then outputs of these output layers 24 can be used directly for control- ling the mode of operation of controller 2. Else a global output layer 25 can be provided whose neu rons indicate, for the robot arm as a whole, wheth er it is operating normally, is in accidental or in deliberate contact. If Q is the number of operating modes, then the neurons of global output layer 25 can be implemented as simple logic gates, e.g. a NOR-gate for normal operation which will output TRUE when none of the sub-networks 22 indicates a contact condition, or OR gates for accidental and deliberate contact, respectively, which will output TRUE whenever one of the sub- networks 22 detects accidental and deliberate contact at its associated link . In practice, the number of neurons Q in the output layers 24 of the sub-networks 22 should be somewhat higher than the number of operating modes, in order to enable each sub-network 22 to provide more in formation to the global output layer 25, e.g. not only whether a contact has occurred or not occurred at an immediately adjacent link but also whether there is a suspicion that a contact might have oc- curred at a remote link, thereby enabling the glob al output layer 25 to draw sounder conclusions.

Reference numerals

1 robot arm

2 controller

3 support

4 end effector

5 link

6 j oint

7 axis

8 axis

9 angular velocity sensor

10 torque sensor

11 neural network

12 FIFO storage

13 screw

14 workpiece

15 hand

1 6 pixel

17 hidden layer

1 8 neuron

1 9 output layer

20 neuron

21 group

22 sub-network

23 hidden layer

24 output layer

25 global output layer