METHOD AND DEVICE FOR TORQUE ESTIMATION

Description

[0001] The present invention relates to methods for estimating a torque acting on a joint of a robot, and to corresponding robots.

[0002] Today, essential tasks in industrial manufacturing are performed by robots that increasingly operate autonomously. Nevertheless, it has become clear that human operators will remain an integral part of modern manufacturing facilities. Development is therefore increasingly focused on the area of ​​safe human-robot collaboration, i.e., creating an environment in which humans and robots can work together without restrictions.

[0003] Particular attention is paid to safety technology, which must be designed to ensure that people and objects are never at risk from the interaction with robots or other autonomously operating units. To achieve this, it is essential that a robot or autonomously operating unit can "recognize" its environment and the actors within it, and take appropriate measures at any time to prevent injury to people or damage to property.

[0004] One way for a robot to perceive its environment is to make it sensitive to contact with objects in its surroundings, thus enabling collision avoidance. For example, a robot can be equipped with torque sensors that can determine the torque acting on its joints. This allows the robot to register even the slightest external forces and implement safe collision protection by reducing its speed, and therefore its kinetic energy, to a level that prevents injury or damage upon unexpected contact.

[0005] Besides directly measuring torque, it is also possible to indirectly derive and estimate the torque acting on a robot using other measured variables. For example, the motor current can be used to infer the applied torque. The advantage over direct measurement using torque sensors is that existing sensors can usually be used, or simpler and less expensive sensors can be employed to determine the measured variables for derivation. However, a disadvantage is that estimation and derivation errors must be considered when determining torque from indirect measurements. These are generally more complex and difficult to determine than measurement errors from a direct measurement.

[0006] DE 10 2012 202 181 A1 discloses a method for determining a torque acting on a link of a robot arm. The determination of the torque is based on determining an input-side and an output-side rotation angle of a gearbox and on a mathematical model of the gearbox, which in particular takes into account the elastic properties of the gearbox.

[0007] US 2015 / 0276436 A1 discloses an angle sensing method by which a control for positioning an output shaft can be implemented using an input shaft encoder and an output shaft encoder with high accuracy and a torsional feedback control, wherein sensing errors of the output shaft encoder are corrected by referring to the input shaft encoder when setting the rotational position information, taking into account the periodicity resulting from the translation.

[0008] DE 10 2010 064 270 A1 and DE 11 2010 004664 B4 each demonstrate further state of the art.

[0009] It is therefore an object of the present invention to provide a method for estimating torque that takes into account the aforementioned disadvantages and allows a precise and reliable determination of torque even for critical applications. In particular, it is an object to provide a precise and reliable determination of torque based on indirect measurement.

[0010] According to one aspect of the present invention, this problem is solved by a method for estimating a torque according to claim 1 and by a corresponding robot according to claim 12.

[0011] The present invention thus aims to determine, by means of an estimation, external influences acting on the moment equilibrium of a robot. This estimation is based on a torsional deformation of the gearbox used.

[0012] This estimate is optimized by compensating for systematic measurement errors and effects inherent to the gearbox. For this purpose, a position-dependent error is determined using a data-driven approach for the case where no external influences are assumed. This position-dependent error is then used to compensate for an actual measurement. The position-dependent error accounts for a systematic measurement error and gearbox disturbances at specific joint positions and is determined beforehand using an identification routine.

[0013] For this purpose, the robot is moved in a predetermined manner during the identification routine to isolate measurement errors for defined joint positions of the robot. The isolated measurement error can then be used to compensate for the actual measurement, for example in the form of a lookup table that provides a compensation value for specific positions or poses of the robot.

[0014] A search table makes it particularly easy and efficient to compensate for an actual measurement by looking up a compensation value for each position in the search table.

[0015] Compensation via a predetermined position-dependent measurement error enables the precise and reliable determination of external influences affecting a robot's torque balance, particularly external torque acting on a joint, solely through torque estimation based on the rotational deformation of the joint mechanism. This eliminates the need for additional torque sensors. Furthermore, it is conceivable that direct measurement using torque sensors could be refined through torque estimation. In this way, redundancy—often required in safety engineering—can be achieved, or the diversity increased, effectively eliminating common-cause errors by employing different measurement methods to determine the torque. The initial problem is thus fully solved.

[0016] In a preferred embodiment, the identification routine includes a defined movement of the joint.

[0017] Preferably, sequential identification is performed with multiple passes, whereby only one joint of the robot is moved in each pass. This allows for the use of a simplified dynamic model of the robot, thus simplifying the isolation of systematic measurement errors. In particular, centripetal and Coriolis effects can be effectively eliminated by appropriately selecting the movement of the individual joint. This design therefore contributes to a particularly simple and precise determination of the position-dependent error.

[0018] A defined motion at a constant velocity is particularly preferred. This also allows moments of inertia to be neglected, further simplifying the robot's dynamic model so that essentially only gravitational moments need to be considered.

[0019] In a further embodiment, a search table is created during the identification routine, which links position data in a joint space of the joint with error rotation angles.

[0020] A search table makes it particularly easy and efficient to compensate for an actual measurement by looking up a compensation value for each position in the search table.

[0021] It is particularly preferred that, for creating the search table, the joint space of the joint is divided into discrete, especially equidistant, sections, and that each section is assigned an aggregated error rotation angle derived from the error rotation angles. This interpolation enables particularly efficient compensation, even under real-time conditions.

[0022] Preferably, the compensated rotation angle in a section is calculated as the sum of the measured rotation angles and the aggregated error rotation angle in that section. The compensation can thus be performed by simple subtraction.

[0023] In a further embodiment, during the identification routine, a tuple is determined at a defined interval. This tuple comprises measured values ​​for the current position and velocity of the joint, as well as the rotation angle of the gear's torsional deformation at the current position. For each tuple, an error rotation angle at the current position is then determined from the measured values. This allows the identification routine to be performed particularly efficiently and quickly.

[0024] In a further embodiment, the gearbox has a position sensor on both the drive and output sides for measuring the angle of rotation of the gearbox's rotational deformation.

[0025] Using position sensors on both the input and output sides of the gearbox, rotational deformation of the gearbox can be determined particularly easily by measuring the offset of the position sensors. This configuration using position sensors is especially cost-effective and thus contributes to a cost-efficient estimation method overall.

[0026] In a preferred embodiment, the transmission is a wave gear with an elastic transmission element.

[0027] Wave gears allow for a particularly good estimation of the torque via the torsional deformation of the gear, since by design the elastic transmission element makes torsional deformation of the gear more pronounced and thus easier to measure.

[0028] According to a further aspect of the present invention, the above-mentioned problem is further solved by a method for estimating a torque acting on a joint of a robot, comprising the steps of: providing a drive; coupling the drive to the joint via a gearbox; measuring a rotation angle of a rotational deformation of the gearbox at a joint position; measuring a motor current of the drive at the joint position; determining a first estimate for the torque at the joint position from the measured rotation angle; determining a second estimate for the torque at the joint position based on the measured motor current; and fusing the first and the second estimates to form a consolidated estimate for the torque at the joint position.

[0029] Alternatively or additionally, the task posed at the beginning can also be accomplished by fusing two indirect measurements. On the one hand, a torque estimate can be made via the motor current, and on the other hand, a torque estimate can be made simultaneously via the rotational deformation, with the results of both estimates then being combined.

[0030] This method also allows for the precise and reliable determination of external influences affecting the robot's torque balance without the need for additional torque sensors. The fusion of measurements advantageously incorporates known characteristics of the respective estimation methods, resulting in a more precise overall outcome than would be possible with a single estimation. Furthermore, merging two indirect measurements provides the redundancy often required in safety engineering and increases diversity by utilizing different measurement techniques. Thus, the aforementioned task is fully accomplished.

[0031] In a preferred embodiment, the merging is carried out with a constant weighting or with time-varying weights.

[0032] By weighting the data, the merger can be adapted to different properties of the estimation methods being merged. This allows for particularly advantageous further optimization of the estimation.

[0033] It is understood that the features mentioned above and those to be explained below can be used not only in the combinations specified, but also in other combinations or on their own, without leaving the scope of the present invention.

[0034] An embodiment of the invention is shown in the drawings and is explained in more detail in the following description. The drawings show: Fig. 1 a schematic representation of a robot according to an embodiment of the present invention, Fig. 2 a flowchart of a process according to a first aspect of the present invention, Fig. 3 a flowchart of a process according to a second aspect of the present invention, and Fig. 4 a schematic representation of a preferred embodiment of moment estimation.

[0035] In the Fig. 1 A robot according to one embodiment of the invention is designated in its entirety by the reference numeral 10.

[0036] In this embodiment, robot 10 is an industrial robot. Industrial robots are universally programmable machines for handling, assembling, or processing workpieces. An industrial robot comprises a manipulator 12 (robot arm), a controller 14, and an effector 16, which can be configured as a tool or gripper.

[0037] The one in Fig. 1 The manipulator 12 shown here has, by way of example, two links 18 and three joints 20. It is understood, however, that the invention is not limited to the number of links and joints shown here. Rather, the presented methods can be applied individually to a large number of joints.

[0038] The links 18 of the manipulator 12 are moved via the joints 20 and driven by a drive 22, which is coupled to the robot 10 via a gearbox 24. For clarity, the gearbox 24 and the drive 22 are shown separately from the manipulator 12. Preferably, however, the drive 22 and gearbox 24 are integrated into the manipulator 12, particularly in the joints 20. Furthermore, each joint 20 can have its own drive 22 and gearbox 24.

[0039] The controller 14 controls the drive(s) 22 of the manipulator 12, enabling it to perform a movement requested by the user. The controller 14 can be a programmable controller, meaning that work and movement sequences can be stored as programs that are executed autonomously by the controller 14.

[0040] Additional sensor information can be acquired by the controller 14 via external or integrated sensors, which influence the operation or movement sequence of the manipulator 12. Typically, a drive-side position sensor 25a is provided, which allows the joint position to be determined. Drive-side here means that the position sensor is located upstream of the gearbox, i.e., on the motor side. Additionally, an output-side position sensor 25b can also be present, i.e., a position sensor located downstream of the gearbox and between the gearbox and the joint. Such a sensor is also referred to as a joint-side position sensor. Using a drive-side and an output-side position sensor, a rotational deformation of the gearbox 24, in the form of a rotation angle, can be easily determined by measuring their relative offsets.

[0041] The controller 14 also executes procedures by which external influences acting on the torque equilibrium of the robot 10 can be determined. In particular, this can be an external force 26 acting on the robot, resulting in a corresponding torque at the joint 20 of the manipulator 12.

[0042] Determining the external torque 26 makes it possible to register even small forces that occur when the robot 10 comes into contact with people or objects. In this way, the robot 10 can be enabled, in the event of a collision with an obstacle, for example, contact with a person, to execute appropriate control functions via the controller 14, causing the robot to retreat or stop its movement.

[0043] In this context, an estimation method is a procedure in which the torque is not determined by direct measurement using torque sensors, but rather derived from another measured quantity. These measured quantities can be acquired, in particular, by sensors that are standard equipment on industrial robots. These sensors can be, among others, position sensors or ammeters, from which relevant measured quantities can be determined directly or indirectly.

[0044] Within the scope of this disclosure, two methods for torque estimation are considered in particular. A first method comprises torque estimation via the motor current and the second method a torque estimation via a rotational deformation of the gearbox.

[0045] To determine the corresponding measured quantity, an ammeter can be used for the first method to measure the current to the drive. The rotational deformation of the gearbox 24 can be determined, among other things, via position sensors 25a, 25b on a drive element and an output element of the gearbox 24. A rotation angle can be determined from the relative offset of the elements to each other, which, similar to the motor current, is in a defined, modelable relationship to the acting torque 26. Such an offset is usually larger in a wave gear due to the elastic element than in other gears. Nevertheless, high measurement accuracy of the position sensors is also crucial here.

[0046] Due to the large gear ratio N of the transmission, the position resolution of the motor-side position sensor is N times higher than the joint-side resolution. The resolution for the measured rotation angle is therefore limited by the joint-side resolution. It is not uncommon for the resolution to be of the same order of magnitude as the expected torsional deformation. Consequently, the influence of a systematic error in the position measurement on the angle measurement is significant. The determination of the acting torque from the rotation angle must therefore be optimized to compensate for this effect.

[0047] To optimize the estimate, the procedures shown below can be carried out individually or in combination. Fig. 2 This shows in a flowchart a method 100 which optimizes an estimate based on a rotational deformation, and Fig. 3 Figure 200 shows a method that combines two estimation methods. The same reference symbols denote the same parts, as in the... Fig. 1 .

[0048] In procedure 100 according to Fig. 2 In a first step, S101 provides a drive 22 of the robot 10 and couples it via a gearbox 24 to at least one joint 20 of the robot 10.

[0049] Then, in step S102, an identification routine is performed.

[0050] The identification routine includes a defined control of the robot 10 in a state in which external influences acting on the moment equilibrium of the robot 10 can be excluded, i.e. the robot can move freely without obstacles in the workspace.

[0051] The aim of the identification routine is to determine a characteristic curve that describes position-dependent errors that are due to systematic measurement errors and position-dependent disturbance effects of the gearbox 24.

[0052] By controlling the robot in a defined manner during the identification routine, a simplified dynamic model of the robot can be assumed, from which the position-dependent error can be determined in the form of a location-dependent error rotation angle. The identification routine can be executed sequentially for this purpose, with only one robot segment moving in each iteration. This eliminates centrifugal and Coriolis effects.

[0053] Furthermore, the defined control can be a movement with constant speed, which allows moments of inertia to be neglected.

[0054] The defined control makes it possible to describe the robot's movement using a simplified dynamic model.

[0055] In general, an industrial robot can be described by a system of motion differential equations: M q q ¨ + C q , q ˙ q ˙ + g q + τ ext = τ J

[0056] Here, M represents the mass inertia matrix, the vector C denotes the generalized constraint moments caused by centripetal and Coriolis forces in the joints, and g is the vector of the generalized gravitational moments. τ J describes the torque transmitted by the gearbox, which results from the motor torque minus the frictional torques of the motor and the gearbox. q(t) further denotes the vector of the motion coordinates of the axes and τ ext represents the external influences that affect the moment equilibrium and which need to be determined.

[0057] The defined control allows individual terms, especially the moments of inertia and constraint moments, to be removed from the model, so that essentially only gravitational moments need to be considered.

[0058] During the identification routine, tuples of measured values ​​are preferably recorded at intervals, each comprising a current position and speed of the joint as well as a rotation angle of the rotational deformation of the gear at the current position.

[0059] Based on the simplified dynamic model, an error can then be assigned to each tuple, resulting in a search table that links position data in a joint space of the joint with error rotation angles. Preferably, the joint space of the joint is divided into discrete, in particular equidistant, sections, and each section is assigned an aggregated error rotation angle from the error rotation angles.

[0060] The data obtained from the identification routine, in particular the search table, can be stored in the controller 14 of the robot 10 or an associated memory and used for compensation.

[0061] It goes without saying that the identification routine must be performed at least once before an actual measurement in order to determine the relevant data. Furthermore, the identification routine can also be performed again at each system startup or at defined intervals to update the position-dependent error.

[0062] Step S103 refers to an actual measurement process in which a rotation angle of a rotational deformation of the gearbox is measured at a joint position.

[0063] The measured rotation angle is then compensated in step S104 by the data determined during the identification routine. Preferably, an error rotation angle at the given position is determined from the data and subtracted from the measured rotation angle to obtain a compensated rotation angle.

[0064] In step S105, an estimate for the acting torque is then given from the compensated rotation angle via a relationship between rotation angle and acting torque that is known in itself.

[0065] The well-known relationship between the angle of rotation and the applied torque can, for example, be modeled by a cubic curve of the following form: τ t , j = k 1 , j Δ q j + k c , j Δ q j 3 ∀ j = 1 , … , n

[0066] This is k l, j for linear stiffness and k c, jfor the cubic stiffness of the j-th member. It is understood that the procedure is not limited to this model, but that other models of the relationship can also be considered.

[0067] According to the in Fig. 2 The torque determined by the described method is significantly more precise than an estimate without error compensation, especially when using gearboxes with a large reduction ratio.

[0068] Alternatively or in addition to the one relating to Fig. 2 The described method allows for a more precise estimate of the moment by fusing two indirect measurements, as can be seen in the following: Fig. 3 is shown.

[0069] Fig. 3 A flowchart shows an alternative method for moment estimation according to the present invention.

[0070] In the alternative method, the torque estimation is optimized by performing two independent estimations and merging their results into an overall result.

[0071] As in the previously described procedure, a drive 22 is first provided in step S201 and coupled to the joint 20 of the robot 10 via a gearbox 24. Subsequently, the two independent estimations are carried out by recording the relevant measured variables.

[0072] In step S202, a rotation angle of a torsional deformation of the gearbox is measured at a joint position. This is preferably done by input and output-side position sensors, the relative offset of which yields a rotation angle that is representative of the torsional deformation of the gearbox 24.

[0073] In step S203, the motor current of the drive is also measured at this joint position, preferably simultaneously.

[0074] Subsequently, in steps S204 and S205, an independent estimate is made for the acting torque.

[0075] In step S204, a first estimate of the torque at the joint position is made from the measured rotation angle. In step S205, a second estimate of the torque at the joint position is made based on the measured motor current.

[0076] Finally, in step S206, the first and second estimates are merged into a consolidated estimate for the torque at the joint position. The aim of this merging is to obtain a better overall result for the torque estimate by combining the individual estimates effectively.

[0077] The merger can be carried out either with a constant weighting or with time-varying weights.

[0078] In the first approach, the fusion during estimation is based on probability theory. For this purpose, the estimates are expressed as probability densities using the following methods: µ t = τ̂ ext,t and µ m = τ̂ ext,m is modeled together with the estimation variants σt 2< and σm 2<. The distributions, if no τ̂ ext is involved, are free of charge for µ t = µ m ≈ 0.

[0079] If the probability density is approximated by a Gaussian distribution, the parameters of the conditional Gaussian probability can be determined according to Bayes' rule, taking into account the two individual estimation probabilities, as follows: τ ^ ext , f = σ m 2 σ t 2 + σ m 2 μ t + σ t 2 σ t 2 + σ m 2 μ m and σ ^ ext , f 2 = σ t 2 σ m 2 σ t 2 + σ m 2 < σ t 2 , σ m 2 .

[0080] This shows that the variance after introducing both estimates is smaller than the lowest individual variance, and the consolidated estimate therefore leads to a better overall result.

[0081] The second approach involves fusing the estimates using time-varying weights. The idea here is to favor the individual estimate that is closer to the expected external moment of 0. This is achieved by adjusting the deviation used to weight the individual estimates based on the square of the estimated external moment.

[0082] This leads to a formulation of the form τ ^ ext , f = τ ^ ext , m 2 τ ^ ext , t 2 + τ ^ ext , m 2 τ ^ ext , t + τ ^ ext , t 2 τ ^ ext , t 2 + τ ^ ext , m 2 τ ^ ext , m , That is, the weights are inversely proportional to the corresponding square distance.

[0083] This reduces the variance, as estimates close to zero are given preferential consideration. Consequently, if, for example, only one estimate increases, the other estimate, which is closer to zero, will dominate the overall merger estimate. Conversely, if both individual estimates increase, for example due to an external factor, the overall estimate will also increase.

[0084] Based on the weighted sum of both estimates, the merged estimate always lies between the two individual estimates, but with a tendency towards an estimate closer to 0.

[0085] It goes without saying that the fusion is not limited to the two approaches mentioned above. Another approach, for example, would be weighting based on velocity, where the torsional moment is favored at lower velocities and the moment determined via the current at higher velocities.

[0086] Fig. 4 Finally, Figure 1 shows a preferred embodiment in which the two methods described above are combined.

[0087] In the exemplary embodiment according to Fig. 4 This results, on the one hand, in an optimized estimate of the gear's rotational deformation by applying compensation for a predefined position-dependent error. On the other hand, the torque estimated in this way is simultaneously fused with a torque estimate based on the motor current.

[0088] In the Fig.4 Above the dashed line, the torque estimate is indicated via the motor current, and below the dashed line, the torque estimate is indicated via the optimized estimate based on the rotational deformation of the gearbox.

[0089] The measured variables present at the input are, on the one hand, the motor current 28 and, on the other hand, the measured rotation angle 30. The measured rotation angle 30 is, as in connection with the Fig. 2 As described in detail, compensation is achieved using a search table 32. The result of the compensation is a compensated rotation angle 34.

[0090] A motor torque 38 is then determined using a motor current model 36. Likewise, a torsional torque 42 is determined using the compensated rotation angle 34 and a rotational deformation model 40 of the gearbox 24.

[0091] A first friction model 44 then takes into account both engine friction and transmission friction, which can be expressed, for example, in the form τ f , m q ˙ = C c , m sgn q ˙ + C v , m q ˙ This example describes C c,m sgn( q̇ ) the Coulomb friction and C v,m q̇ The viscous friction of the transmission. Taking both friction components into account, the actual transmitted torque 46 can be determined from the motor torque. It is understood that the method is not limited to the friction model 44 presented here, but that other models can also be considered.

[0092] Similarly, the transmitted torsional moment 48 can be determined from the torsional moment 42 minus the gear friction, which can be determined using a second friction model 50. The second friction model 50 can, for example, advantageously include only the gear friction: τ f,t ( q̇ ) = C v,t q̇ Here too, other models of friction are conceivable.

[0093] From the transmitted motor torque 46 and the transmitted torsional torque 48, a first estimate 54 and a second estimate 56 for the external torque 26 at the present position can then be determined in a manner known per se.

[0094] While this determination is possible directly via the robot's dynamic model, it is advantageous to determine the values ​​indirectly using a perturbation observer 52, since this requires neither directly measuring the acceleration nor calculating the inverse mass inertia matrix. The calculation can thus be advantageously simplified.

[0095] Finally, the first and second estimates 54, 56 are merged into a consolidated estimate 58. The merger can be described in relation to Fig. 3 as explained above, and encompass the various approaches to merging.

Claims

1. Method (100) for estimating a torque acting on a joint (20) of a robot (10), comprising the steps of: - providing (S101; S201) a drive (22); - coupling (S101; S201) the drive (22) to the joint (20) via a transmission (24); - measuring (S103; S202) a rotational angle (30) of a rotational deformation of the transmission (24) at a joint position; characterized by: - performing (S102) an identification routine for determining a position-dependent error rotational angle of a rotational deformation of the transmission (24), wherein during the identification routine a lookup table (32) is generated which associates position data in a joint space of the joint with identified error rotational angles; - compensating (S104) the measured rotational angle (30) by means of the position-dependent error rotational angle from the lookup table (32); and - determining (S105) a first estimate (54) of the torque at the joint position from the compensated rotational angle (34), wherein the robot (10) is controlled in a defined manner during the identification routine, and the robot (10) is in a state during said defined control in which external influences acting on a moment equilibrium of the robot (10) can be excluded.

2. The method according to claim 1, wherein the identification routine comprises a defined movement of the joint (20).

3. The method according to claim 2, wherein the defined movement is a movement at a constant velocity.

4. The method according to any one of claims 1 to 3, wherein, for generating the lookup table (32), the joint space of the joint (20) is divided into discrete, in particular equidistant, sections, and an aggregated error rotational angle derived from the identified error rotational angles is assigned to each section.

5. The method according to claim 4, wherein the compensated rotational angle (34) in a section is the measured rotational angle (30) of the aggregated error rotational angle in this section.

6. The method according to any one of the preceding claims, wherein, during the identification routine, within a defined interval, a tuple is determined which comprises measurement values for a current position and velocity of the joint (20) as well as a rotational angle of the rotational deformation of the transmission at the current position, and for each tuple an error rotational angle at the current position is determined from the measurement values.

7. The method according to any one of the preceding claims, wherein the transmission (24) comprises a position sensor on a drive side and on an output side for measuring the rotational angle of the rotational deformation of the transmission.

8. The method according to any one of the preceding claims, wherein the transmission (24) is a harmonic drive having an elastic transmission element.

9. The method according to any one of the preceding claims, further comprising: - measuring (S203) a motor current (28) of the drive (22) at the joint position; - determining (S205) a second estimate (56) of the torque at the joint position based on the measured motor current (28); and - fusing (S206) the first and second estimates (54, 56) into a consolidated estimate (58) of the torque at the joint position.

10. The method according to claim 9, wherein the fusing is performed with a constant weighting or with time-varying weights.

11. A computer program comprising instructions which, when executed by a computer, cause the computer to carry out the method according to any one of claims 1 to 10.

12. A robot comprising: a joint (20), a drive (22), a transmission (24) for coupling the drive (22) to the joint (20), and a controller (14) for estimating a torque acting on the joint (20), wherein the controller (14) is configured to, measure a rotational angle (30) of a rotational deformation of the transmission (24) at a joint position, characterized in that the controller (14) is configured to: perform an identification routine for determining a position-dependent error rotational angle of a rotational deformation of the transmission (24), wherein performing the identification routine includes generating a lookup table (32) that associates position data in a joint space of the joint with identified error rotational angles, compensate the measured rotational angle (30) by means of the position-dependent error rotational angle from the lookup table (32), and determine a first estimate (54) of the torque at the joint position from the compensated rotational angle (34), wherein the robot (10) is controlled in a defined manner during the identification routine, and the robot (10) is in a state during said defined control in which external influences acting on a moment equilibrium of the robot (10) can be excluded.

13. The method according to any one of claims 1 to 10 or the robot according to claim 12, wherein the robot (10) moves freely of obstacles within a workspace during the defined control.