Hydraulic Drive System

Abstract

To provide a hydraulic driving system for remotely controlling a manipulator capable of accurately estimating 3D external force applied to an end effector of the manipulator without installing a sensor on the manipulator itself.SOLUTION: A hydraulic driving system including a mobile manipulator is formed of at least one joint for rotating a first connection part on an X-Z plane, at least one joint for rotating a second connection part on another plane, and an end effector unit. A 6-axis force torque sensor for learning data is installed in the end effector unit. The hydraulic driving system has an interface and a 3D force estimation unit for commanding movement of a robot. An input of the 3D force estimation unit is constituted by the number of pressure sensors and the number of pieces of attitude data. An output of the 3D force estimation unit is constituted by three outputs of force on the respective axes, and is calculated from a neural network model obtained by supervised learning.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a method for estimating 3D external forces applied to an end effector of a hydraulically driven manipulator system, and more particularly to a method for estimating forces in any possible manipulator pose. [Background technology]

[0002] Manipulating heavy or long objects in extreme environments such as highly radioactive buildings or the deep sea requires large, safe and controllable powered machines. Hydraulic-driven manipulators are considered suitable for these types of applications due to their characteristics of high power and durability.

[0003] However, in these harsh environments, in order to achieve higher precision controllability of hydraulically driven manipulators, it is not possible to directly install force sensors on the end effectors of the manipulators.

[0004] Therefore, a method for estimating the 3D external forces applied to the end effector of a manipulator is needed.

[0005] Background art in this technical field includes, for example, a technique such as that disclosed in Patent Document 1. Patent Document 1 proposes a method for estimating a force vector of a load using a simple configuration or a manipulator and one torque sensor at each joint.

[0006] In Patent Document 1, the manipulator is set to a specific posture while holding a load. In that state, the torque at each joint is measured without measuring the posture. From the dynamic relationship of the manipulator, the load weight and the horizontal position of the center of gravity of the load can be calculated.

[0007] Furthermore, Patent Document 2 proposes a method for estimating the bucket attitude using a hydraulic drive system in order to improve the reliability of the automatic excavation operation of an excavator, which is caused by sensor failure due to a large impact at the tip of the excavator.

[0008] In the '2006 patent, pressure and attitude data of the boom and arm, as well as pressure data of the bucket are used for estimation. Although the '2006 invention is not intentionally used to estimate the weight of the load, the proposed configuration makes it possible to estimate the force applied to the tip, especially for 2D forces in the xy plane of the described axes. [Prior art documents] [Patent documents]

[0009] [Patent Document 1] JP 2020-151812 A [Patent Document 2] JP 2019-167731 A Summary of the Invention [Problem to be solved by the invention]

[0010] However, the method of Patent Document 1 cannot be used to estimate external forces in an undetermined posture, since the calculation unit does not use posture data.

[0011] Also, this method can only be used with electric motor systems since hydraulically driven manipulators cannot be set to equal positions using constant torque or pressure control due to system hysteresis.

[0012] Furthermore, the method of Patent Document 2 lacks information on the bucket attitude, base attitude, and base pressure or torque, making it impossible to estimate the 3D external force at the tip of the bucket.

[0013] Therefore, an object of the present invention is to provide a highly reliable hydraulic drive system for remotely operating a manipulator, which is capable of estimating with high accuracy the 3D external forces acting on the end effector of a manipulator in any manipulator posture, without installing sensors on the manipulator itself. [Means for solving the problem]

[0014] In order to solve the above problems, the present invention provides a hydraulic drive system equipped with a mobile manipulator that includes a crawler and a manipulator. In The manipulator includes at least one joint for rotating a first joint in an xz plane and at least one joint for rotating a second joint in another plane; End effector It is composed of The end effector a 6-axis force torque sensor used only for obtaining ground truth for learning data is installed in the system, the joints are driven by at least one fluid actuator that moves a piston by supplying fluid, the joints include a fluid driver that supplies and discharges fluid from a fluid reservoir to the fluid actuator, a pump that supplies fluid from the fluid reservoir to the system, a relief valve that adjusts the pressure of the fluid supplied from the pump to the system, one solenoid valve that controls the flow of fluid from the system to the fluid actuator, and another solenoid valve that controls the flow in the reverse direction, a pressure unit that measures the fluid pressure at each joint using at least one pressure sensor, a posture unit that calculates joint position or angle data using at least one flow measurement element, a controller that generates control signals for the valves, an interface that displays the state of the robot and commands the movement of the robot, and a 3D force estimation unit, the input of the 3D force estimation unit being configured by the number of pressure sensors and the number of posture data, the output of the 3D force estimation unit being configured by three outputs of forces on each axis, and the output of the 3D force estimation unit being configured by three outputs of forces on each axis. The end effector or gripper position, and the output of the 3D force estimation unit is calculated from a neural network model obtained by supervised learning. A hydraulic drive system further comprising a motor for electromechanically controlling the relief valve, the pressure sensor is installed at the outlet of the relief valve, and data from the pressure sensor is included in the input of the 3D force estimation unit, the pressure unit installed at one joint is obtained using two pressure sensors that measure the pressure state of the lower chamber and the rod chamber of the fluid actuator, and the attitude unit at one joint is composed of two flow rate measuring elements, and calculates the attitude by subtracting the flow rate measured by one flow rate measuring element on the bottom surface of the fluid actuator and one flow rate measuring element on the rod side. It is characterized by: Effect of the Invention

[0015] According to the present invention, in a hydraulic drive system for remotely operating a manipulator, it is possible to realize a highly reliable hydraulic drive system that can estimate with high accuracy the 3D external forces acting on the end effector of the manipulator in any manipulator posture, without installing sensors on the manipulator itself.

[0016] This makes it possible to prevent damage to the object being worked on or the manipulator itself, even in extreme environments such as highly radioactive buildings or the deep sea.

[0017] Problems, configurations and effects other than those described above will become apparent from the following description of the embodiments. [Brief description of the drawings]

[0018] [Figure 1] 1 is a diagram showing a schematic configuration of a hydraulic drive system according to a first embodiment of the present invention. [Diagram 2] FIG. 1 shows the basic configuration of one actuator having one pressure sensor and one flow rate measuring element. [Diagram 3] 2 is a diagram illustrating the configuration and functions of a 3D force estimation unit in FIG. 1. [Figure 4] FIG. 1 shows a configuration of one actuator having two pressure sensors and two flow measuring elements and one pressure sensor in a relief valve. [Diagram 5] FIG. 13 is a diagram showing a method for successively learning an estimator model in three stages using three types of data sets having different forms of external forces. [Figure 6] FIG. 11 is a diagram showing a schematic configuration of a hydraulic drive system according to a second embodiment of the present invention. [Figure 7] FIG. 7 is a diagram illustrating the configuration and functions of a 3D force estimation unit in FIG. 6. [Figure 8] FIG. 1 illustrates the configuration of moment lengths during various types of applied forces. [Figure 9] 13 is a flowchart showing a procedure for calculating moment lengths without prior information. [Figure 10] FIG. 11 is a diagram showing a configuration of a continuum joint according to a third embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0019] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In the drawings, the same components are designated by the same reference numerals, and detailed description of overlapping parts will be omitted.

[0020] Moreover, the following description is an example of the proposed embodiment, and is not intended to limit the scope of the claims. Furthermore, the working principle of the estimation system is an example provided to enable a person skilled in the art to implement the method according to the present invention, and is not intended to limit the scope of the claims. EXAMPLES

[0021] A hydraulic drive system and a 3D external force estimation method according to a first embodiment of the present invention will be described with reference to Figs.

[0022] The main components of the hydraulic drive system of this embodiment are shown in Figures 1 and 2. As shown in Figure 1, the hydraulic drive system of this embodiment is composed of a fluid-driven mobile manipulator 1 connected to a long hose 10 controlled by a fluid driver 2. Since the mobile manipulator 1 operates in a harsh environment, only the robot (i.e., the main body of the mobile manipulator 1) and the hose 10 are exposed to the environment, while the hydraulic drive system 8 and interface 7 are kept in a safer location.

[0023] 2, typically, fluid is supplied from a fluid reservoir 25 and discharged to the cylinder actuator 12 by using a motor pump (electric pump) 24. The pressure of the fluid is regulated by a relief valve 23 and supplied to the lower chamber 13 of the cylinder actuator 12 by controlling the flow of fluid through a solenoid valve 21.

[0024] The condition of the fluid flowing in the cylinder actuator 12 and along the hose 10 is measured by a pressure sensor 31 inside the pressure unit 3 and a flow measuring element 41 inside the attitude unit 4. By pushing or pulling the piston 15, it is mechanically translated into a manipulator joint angle and the attitude of each joint of the mobile manipulator 1 is calculated from the measured flow entering the lower chamber 13.

[0025] The calculated poses and pressures of the mobile manipulator 1 are input to a 3D force estimation unit 5, which outputs as manipulator end effector forces on three axes, Fx, Fy, and Fz. These three outputs are then displayed on an interface 7 or become force feedback parameters for the controller 6, which actuates the fluid driver 2 to further control the movement of the manipulator.

[0026] The number of inputs to the 3D force estimation unit 5 is defined by the number of output vectors of the pressure unit 3 and the number of output vectors of the attitude unit 4. Assume that the manipulator is composed of five joints, three of which rotate in the xz plane, one joint rotates in the xy plane, and the other joint rotates in the yz plane. If each joint is driven by one cylinder actuator 12, there is a 5-cylinder actuator whose state is measured by five pressure sensors 31 and five flow measurement elements 41. Therefore, the inputs to the 3D force estimation unit 5 are 10 vectors.

[0027] However, the number of actuators in one joint is not limited to one cylinder actuator, and it is possible to have two or more cylinder actuators for rotating one joint. Therefore, if there are two cylinder actuators for one joint, the number of inputs to the 3D force estimation unit 5 is 20 vectors.

[0028] It should be noted that the flow measurement element 41 is not limited to a conventional flow sensor, but may be any other sensor capable of measuring either the volume of fluid movement or the speed of fluid movement. An example of such a sensor is a cylinder with a position sensor attached to the piston, in which case the volume of fluid can be calculated from the displacement of the piston measured by the position sensor and the diameter of the cylinder. The volume displacement of the sensor can be converted into the displacement of the cylinder actuator and calculated as the attitude state.

[0029] 3 shows the inside of the 3D force estimation unit 5, which is a neural network system. The input vectors of the pressure state from the pressure unit 3 and the posture state from the posture unit 4 are input to a pre-processing calculation unit 51. The pre-processing calculation unit 51 is composed of at least rate synchronization and input normalization from different frequencies of the sensor units.

[0030] The normalized inputs are input to nodes on the neural network's input layer 52. The input layer 52 is connected to nodes on the neural network's hidden layer 53 and through nodes on the neural network's output layer 54. The number of nodes in the hidden layer 53 is at least as many as the number of nodes in the input layer 52.

[0031] The hidden layer 53 is not limited to only a fully connected (FC) type node or a single layer, but can be a multi-layer of RNN type, LSTM type, GRU type, or FC type, since it is a fluid-driven actuator that inherits hysteresis characteristics. The number of nodes on the output layer 54 is three. The output from each node of the output layer 54 is post-processed in the post-processing calculation unit 55. The post-processing calculation unit 55 consists of at least denormalization of the output.

[0032] To train the 3D force estimation unit 5 , the end effector position of the manipulator is controlled by the operator from an interface 7 via a controller 6 .

[0033] The controller 6 further controls the valves of each joint, which generates a pushing or pulling force on the cylinder actuator 12. This pushing or pulling force changes the attitude state of the manipulator and the pressure state inside the actuator. Forces are applied to the end effector by the weight of the load or by external forces from contact with the environment.

[0034] The load also includes a zero-weight object that is the basis of the 3D force estimation unit 5. This external force is measured using a 6-axis force torque sensor 11. A learning data set for the 3D force estimation unit 5 can be obtained by randomly controlling the joints of the manipulator and logging the pressure state, attitude state, and force torque sensor data.

[0035] The weighting for the connection of each node is trained using a machine learning method. For example, when the neural network model is trained using a supervised learning method, the correct force applied to the end effector of the manipulator is measured by the 6-axis force torque sensor 11 with the input of the given pressure and posture state, and the weighting for the connection of each node is calculated using a backpropagation method or the like. The 6-axis force torque sensor 11 is used only for training the model and is not used for the actual estimation.

[0036] To estimate the 3D external forces applied to the end effector of the manipulator, the controller 6 generates control motions for the manipulator, and the pressure and attitude states are measured by the corresponding pressure unit 3 and attitude unit 4, respectively, and these states are input as input signals to the 3D force estimation unit 5.

[0037] The input signals input to the 3D force estimation unit 5 are further processed by trained nodes 52, 53, 54 which result in estimated forces to be applied on each axis in local coordinates equivalent to the 6-axis force torque sensor 11. The output forces from the local coordinate sensors can then be transformed into the desired coordinates and that transformation placed within the controller 6.

[0038] With this main configuration, it is possible to realize a force estimation system capable of estimating the 3D force acting on the end effector of the fluid pressure-driven manipulator by using the pressure data and posture data as input to the 3D force estimation unit 5.

[0039] In addition, the force estimation system of this embodiment is configured to include additional data of the fluid state as shown in Figure 4 in order to improve the estimation accuracy. In this figure, the number of cylinder actuators 12, 12a depends on the number of joints of the robot and its basic configuration. Because there are multiple actuators, without knowing the base of the adjusted pressure, the estimation accuracy is limited by the setting of the relief valve 23.

[0040] In this embodiment, an electromechanical relief valve is used for the relief valve 23, and the pressure source for actuation is automatically adjusted by the controller 6 or by a command from the interface 7. A pressure sensor 32 is also installed at the output of the relief valve 23, so that the adjusted pressure can be measured. By having a controllable relief valve 23, it is possible to estimate the pressure in the rod chamber 14 or 14a of each cylinder, so that the pressure difference in both the lower chamber 13 and the rod chamber 14, or the lower chamber 13a and the rod chamber 14a can be counted during training of the estimation model.

[0041] In order to further improve the estimation accuracy, a pressure sensor is added to the rod chamber 14 side of the cylinder actuator 12, which provides a more accurate pressure state of the cylinder actuator 12 when estimating the force during the movement of the manipulator caused by some change in the solenoid valve 22. In addition, the estimation accuracy is further improved by making the attitude state of the manipulator more accurate due to the change in the position of the piston 15 controlled by the solenoid valve 22. The attitude is calculated based on the difference in the fluid flow between the lower chamber 13 side and the rod chamber 14 side.

[0042] Knowing that the hydraulic actuator has hysteresis characteristics, the up and down movement of the manipulator or the push and pull movement of the piston 15 may cause pressure differences even at the same piston 15 position, therefore the actual posture itself is not calculated by the posture unit 4, but the sensing result of the flow measurement element is used as a direct input to the 3D force estimation unit 5. Therefore, the hysteresis of the influence that causes errors in the posture unit 4 is directly trained into the network, improving the estimation accuracy.

[0043] Generally, any dataset trained on the estimation system is aggregated in sum, and the model learning process obtains training data in a random manner from the aforementioned aggregated dataset. However, the force vectors caused by loads and external contact forces have different characteristics. The result of the force vector caused by the load is obviously a vector in the direction of gravity, while the external contact forces are usually oriented in a random direction perpendicular to the contact area. If all datasets are directly aggregated and trained, the 3D force vectors in the training data are inconsistent, which leads to overfitting the learning process.

[0044] The procedure for training an estimation system for force vector direction discrepancies is shown in Figure 5. In this figure, the dataset is shown as having three types of sources of external forces applied to the end effector: force on the load 501, contact force with the environment 503, and contact force with the environment while holding the load 505.

[0045] Training of the estimation model is performed in three stages based on the type of dataset: The first training stage 502 is a stage where only the dataset of external forces caused by the weight of the load 501 is trained. By doing this, the model can correctly learn the characteristics of the 3D forces in any manipulator pose that may be caused by the load.

[0046] The model trained in the first training stage 502 is then used as an initial model for a second training stage 504, trained using a second type of data set 503, which is a contact force with the environment (disturbance force). Similarly, the model trained in the second training stage 504 is used as an initial model for a third training stage 506, trained using a third type of data set 505, which is a load weight and disturbance force.

[0047] In addition, although Fig. 5 shows three types of external force sources, the present invention is not limited to these three types. In addition, a static position of the end effector can be set and contact force variations can be added by applying attitude control changes to the actuator. This allows the dynamic generation of pressure conditions and flow measurements that cause external forces. In addition, the order of the training stages is not limited to Fig. 5 and different orders of data types, including other variations of data set types, are possible.

[0048] Since the robot operates in a harsh environment, severe temperature changes are inevitable. To cover the estimation range from laboratory temperature conditions to wider temperature changes, an additional temperature sensor (not shown) is used as an input of the 3D force estimation unit 5. Since the temperature of each joint is generally equal, one temperature sensor is added to the output of the relief valve 23 upstream of the solenoid valve 21 of each joint. Since the temperature change of the fluid also affects the change of the pressure state, the addition of a temperature sensor can improve the estimation ability under different temperature conditions.

[0049] To train a dataset under different temperature conditions, it is not necessary to change the entire environment temperature, but at least the fluid temperature is changed by some mechanism that controls the fluid temperature. The temperature range depends on the fluid medium, which can be water, oil, or gas.

[0050] Training can also be performed by varying the environmental temperature. By doing this, the 6-axis force-torque sensor 11 must also be calibrated, since temperature changes lead to measurement offsets for the sensor. Typically, the force sensor is calibrated while removed from the manipulator and placed flat on the ground. This is a training procedure that is inconvenient for the operator. Therefore, an additional calibration procedure is required.

[0051] In this embodiment, the six-axis force torque sensor 11 is calibrated while attached to the end effector without removing it. Instead of placing the force sensor flat on the ground, the six-axis force torque sensor 11 can be calibrated while setting the manipulator to any predetermined posture. This method can simplify the calibration procedure. Training is performed by referring to the data of the six-axis force torque sensor 11 at those predetermined postures while having that posture as the zero state of the force sensor.

[0052] The training process can be performed normally with that calibration method. However, during the estimation process, it is important to transform the results of the estimation by the 3D force estimation unit 5 from a local predefined initial calibration pose to the desired coordinates. Also, the training data set for each stage includes this calibration data so as to have a standard zero force of the measurement system. Furthermore, the training stage, while including this calibration data of the force sensor, can add a range of temperatures to improve the estimation accuracy. EXAMPLES

[0053] Second Embodiment A hydraulic drive system and a 3D external force estimation method according to a second embodiment of the present invention will be described with reference to FIGS.

[0054] FIG. 6 is a diagram showing a schematic configuration of the hydraulic drive system of this embodiment. As shown in FIG. 6, the hydraulic drive system of this embodiment has a configuration similar to the basic configuration of the first embodiment, but the 3D force estimation unit 50 has three outputs of force and three outputs of torque. The outputs of this 3D force estimation unit 50 are equal as if the 6-axis force torque sensor 11 was attached to the system. In addition, the moment length of the acting torque is added to the input of the 3D force estimation unit 50, while this moment length is calculated by the moment length calculation unit 61 inside the controller 6. The motion control unit 60 is a control unit that converts the input interface 7 or the autonomous motion generation controller into a solenoid valve opening and closing command.

[0055] The detailed process of the 3D force estimation unit 50 is shown in Figure 7. The working principle of the training and estimation process is similar to that of the first embodiment, but the input of the pre-processing calculation unit 511 is added with the moment length r, while this value is set as the center of gravity position of the end effector. Therefore, the number of nodes on the hidden layer 531 is at least as many as the number of nodes in the input layer 521, while the input layer 521 is also added with one node. And the output layer 541 has six nodes, and the result of the neural network calculation is further processed by the post-processing calculation unit 551 to output three estimated forces on each axis and three estimated torques on each axis.

[0056] This basic configuration of Example 2 allows for estimation of not only the applied force but also the torque applied to the manipulator's end effector.

[0057] During the training process, the moment length is known because the operator sets the weight of the load or the contact position with the environment. However, in the real-time estimation process, the moment length needs to be calculated. The moment length calculation unit 61 gives a constant output by default depending on the value of the center of gravity position 603 of the end effector. In addition, the actual torque error (either rmse, mse, or mae) from the measurement value of the 6-axis force torque sensor 11 is added to the metric for training the neural network, and the force estimation result is multiplied by the moment length of the center of gravity position 603 of the end effector.

[0058] To improve the accuracy of the moment length calculation, a center of gravity equation for multiple rigid bodies can be used, as shown in the following equation (1): The sum of the product of the end effector mass and the center of gravity of the end effector and the product of the load mass and the end effector length is divided by the sum of the end effector mass and the load mass.

[0059]

number

[0060] Here, me is the mass of the end effector, re is the center of gravity of the end effector, m0 is the estimated mass of the load (or external force), and r0 is the center of gravity of the load (or external force).

[0061] FIG. 8 shows the location of the center of gravity from multiple rigid bodies. FIG. 8(a) shows the external force caused by only the weight of the end effector, where the location (or moment length) 111 of the center of gravity of the end effector and the magnitude of the weight (force) of the end effector are known to be constant. FIG. 8(b) shows the external force caused by the weight of a load, where the location of the load's center of gravity is shown as 112 and the locations of the centers of gravity of both the end effector and the load as 113. FIG. 8(c) shows the external force caused by contact with the environment 114, and FIG. 8(d) shows the external force caused by contact with the environment 114 while holding an object (load) 110.

[0062] In this embodiment, since it is not possible to install any kind of sensor at the position of the end effector, the exact position of the external force is unknown. However, by considering ideal conditions, the position of the external force can be considered to be equal to the moment length of the total force.

[0063] During training, the load mass and the applied forces are known in advance, however, for real-time estimation this information needs to be obtained from some calculation.

[0064] 9 shows the procedure for calculating the moment length without prior information. In principle, this procedure is performed as part of the controller 6. A check block 601 determines whether the moment length has been calculated. If the moment length has been calculated (YES), the 3D force estimation unit 50 calculates the estimated forces as usual by using the calculated moment length. On the other hand, if the moment length has not been calculated (NO), the controller 6 commands the manipulator to proceed to the pose calibration block 602, which will be described later.

[0065] The system then estimates the force applied to the end effector by using the 3D force estimation unit 50, whose input moment length is the center of gravity position 603 of the end effector. The estimated Fz, which is the output data 650 of the 3D force estimation unit 50, is used for calculation in the moment length calculation unit 61. Finally, the moment length position 611 calculated by the moment length calculation unit 61 is input to the 3D force estimation unit 50 and utilized for real-time estimation. This output data 651 of the 3D force estimation unit is used by the controller 6 to execute other control procedures. The calculation in the moment length calculation unit 61 can be expressed by the following equation (2):

[0066]

number

[0067] Here, Fz is the output data 650 (estimation result) of the 3D force estimation unit using the calibration pose determined in the pose calibration block 602. The calibration pose is any predefined joint pose where the end effector orientation is pointing up or down perpendicular to the global z-axis while the end effector position is kept constant by a single or several poses.

[0068] In this pose, the center of mass of the end effector in the global x-axis is equal to the center of mass of the load in the global x-axis. The center of mass location is different in the z-axis, but this difference is not important since we are only measuring the sum of the forces. This given pose allows us to improve the accuracy of the estimated force Fz while still using the moment length value as the center of mass location of the end effector.

[0069] As in the first embodiment, in this embodiment, the accuracy of the estimator can be improved by adding an electromechanical relief valve 23 so that the pressure source for actuation can be adjusted by command from the controller 6 or the interface 7. Also, a pressure sensor 32 is added to the output of the relief valve 23 so that the adjusted pressure can be measured. When estimating the force during the movement of the manipulator caused by the change in the solenoid valve 22, the accuracy of the estimator can be further improved by adding a pressure sensor 33 to the rod chamber 14 side of the cylinder actuator 12. Also, by adding a flow measurement element 42 to the rod chamber 14 side, all flow measurement elements are directly used for the input to the 3D force estimator unit 50, and the accuracy of the estimator can be further improved.

[0070] Also, in the present embodiment, similarly to the first embodiment, the estimable range of the estimator can be expanded by installing a temperature sensor (not shown) at the output of the relief valve 23. To further improve the estimable range, three or more types of learning data sets are trained successively. Also, the calibration of the force sensor can be made simpler by using a six-axis force-torque sensor 11 attached to the manipulator, and the calibration is performed by moving the manipulator to a predetermined calibration posture. EXAMPLES

[0071] Third Embodiment With reference to FIG. 10, a hydraulic drive system and a 3D external force estimation method according to a third embodiment of the present invention will be described.

[0072] 10 is a diagram showing the configuration of the continuum joint of this embodiment. The hydraulic drive system 8 of the embodiment 3 has the same basic configuration as that of the embodiment 1. In this embodiment, the mobile manipulator itself is configured not only with a general rotational joint, but also with at least one additional continuum joint having flexible movement in 3D space.

[0073] In Figure 10, the structure of the continuum joint is shown using top and side views. From top to bottom, the one-unit continuum joint is composed of four pairs of cylinder actuators 121, 122, 123, 124, and each pair of actuators is connected to each other at the rod side and the bottom side. The top 151 of the one-unit continuum joint is a plate to which the four pairs of cylinder actuators 121, 122, 123, 124 are attached, and the bottom 152 of the unit is also a plate to which another four pairs of cylinder actuators 125, 126, 127, 128 (reference numbers 127 and 128 are not shown) are attached.

[0074] The input ports of these actuators are connected in pairs, so that one solenoid valve can control two cylinders simultaneously. Also, a ball joint (not shown) is attached between the two rods, so that the pair of actuators can move by expanding, contracting, and bending.

[0075] If the pressure state of the cylinder actuator 124 and its paired cylinder actuator 128 (not shown) is made equal to that of the cylinder actuator 121, and the pressure state of the paired cylinder actuator 125 is greater than the pressure states of the cylinder pairs 122, 126 and 123, 127 (not shown), the continuum joint will bend on the y-axis. Other combinations can also produce a similar mechanism. If all pressure states are equal, the actuators will expand or contract on the z-axis depending on the previous pressure state. If only one pair of cylinder actuators 124, 128 has a pressure state greater than the other pairs of cylinder actuators, the continuum joint will bend on the xy-axis.

[0076] Also, this embodiment is not limited to a single unit of continuum joint, but multiple units of continuum joints can be stacked with inputs parallel to all units as shown in Fig. 10. Using this configuration, flexible motions that can move in the xyz plane can be generated, and the forces applied to each axis can also be estimated with fewer actuating parts and fewer pressure and attitude measurement units.

[0077] As in the first embodiment, in this embodiment, the accuracy of the estimator can be improved by adding a relief valve 23. Also, a pressure sensor 32 is added to the output of the relief valve 23, and the adjusted pressure can be measured. Also, the accuracy of the estimator can be further improved by adding a pressure sensor 33 to the rod chamber 14 side of the cylinder actuator 12. Also, by adding a flow measurement element 42 to the rod chamber 14 side, all flow measurement elements are directly used for the input to the 3D force estimator unit 50, and the accuracy of the estimator can be further improved.

[0078] Also, in the present embodiment, similarly to the first embodiment, the estimable range of the estimator can be expanded by installing a temperature sensor (not shown) at the output of the relief valve 23. To further improve the estimable range, three or more types of learning data sets are trained successively. Also, the calibration of the force sensor can be made simpler by using a six-axis force-torque sensor 11 attached to the manipulator, and the calibration is performed by moving the manipulator to a predetermined calibration posture. EXAMPLES

[0079] Fourth Embodiment A hydraulic drive system and a 3D external force estimation method according to a fourth embodiment of the present invention will be described.

[0080] The basic configuration of the fourth embodiment is similar to that of the third embodiment, but the 3D force estimation unit 50 has three outputs of force and three outputs of torque. This configuration allows estimation of these six outputs even when using a continuum joint with a small number of pressure units and attitude units.

[0081] As in the third embodiment, in this embodiment, the accuracy of the estimator can be improved by adding a relief valve 23. Also, a pressure sensor 32 is added to the output of the relief valve 23, and the adjusted pressure can be measured. Also, the accuracy of the estimator can be further improved by adding a pressure sensor 33 to the rod chamber 14 side of the cylinder actuator 12. Also, by adding a flow measurement element 42 to the rod chamber 14 side, all flow measurement elements are directly used for the input to the 3D force estimator unit 50, and the accuracy of the estimator can be further improved.

[0082] Also, in the present embodiment, similarly to the third embodiment, the estimable range of the estimator can be expanded by installing a temperature sensor (not shown) at the output of the relief valve 23. To further improve the estimable range, three or more types of learning data sets are trained successively. Also, the calibration of the force sensor can be made simpler by using a six-axis force-torque sensor 11 attached to the manipulator, and the calibration is performed by moving the manipulator to a predetermined calibration posture.

[0083] The present invention is not limited to the above-described embodiments, and includes various modified examples. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to those including all of the configurations described. It is also possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. It is also possible to add, delete, or replace a part of the configuration of each embodiment with another configuration. [Explanation of symbols]

[0084] 1...Movement manipulator 2...Fluid driver 3…Pressure unit 4...Attitude unit 5,50…3D force estimation unit 6…Controller 7. Interface 8...Hydraulic drive system 10,10a…Hose 11...6-axis force torque sensor 12, 12a...Cylinder actuator 13, 13a...Lower chamber 14, 14a…Rod chamber 15,15a…Piston 21, 21a, 22, 22a...Solenoid valves 23...Relief valve 24...Motor pump (electric pump) 25…Fluid reservoir 31, 31a, 32, 33, 33a...Pressure sensor 41, 41a, 42, 42a...flow rate measuring element 51,511…Pre-processing section 52,521…Input layer of neural network 53,531…hidden layer of neural network 54,541…output layer of neural network 55,551…Post-processing calculation section 60...Motion control unit 61...Moment length calculation unit 110…Object (load) 111…Moment length of end effector 112...Center of gravity of load 113...The position of the center of gravity of both the end effector and the load 114...Environment 121, 122, 123, 124, 125, 126, 129, 130, 133, 134...Cylinder actuators (for continuum joints) 141, 142...Fluid inlets to the rod chamber of the cylinder 145, 146...Fluid inlets to the lower chamber of the cylinder 151, 153…Apex of continuum joint 152, 154...Bottom of continuum joint 501...Force against load (weight of load) 502...First training stage of the estimation model 503…Contact force with the environment (disturbance force) 504...Second training stage of the estimation model 505...Load weight and disturbance force 506…Third training stage of the estimation model 601...Confirmation block 602...Attitude calibration block 603: Center of gravity of end effector 604...Termination confirmation control block 611…Moment length position 650,651…Output data of 3D force estimation unit

Claims

1. A hydraulic drive system equipped with a mobile manipulator consisting of a crawler and a manipulator, The manipulator includes at least one joint for rotating a first joint in an xz plane, at least one joint for rotating a second joint in another plane, and an end effector; A six-axis force torque sensor is installed on the end effector and is used only to obtain ground truth for learning data; the joint is actuated by at least one fluid actuator that moves a piston by supplying a fluid; a fluid driver for supplying and discharging fluid from a fluid reservoir to the fluid actuator; a pump for supplying fluid to the system from the fluid reservoir; a relief valve for adjusting the pressure of the fluid supplied from the pump to the system; one solenoid valve for controlling the flow of fluid from the system to the fluid actuator and another solenoid valve for controlling reverse flow; a pressure unit for measuring the fluid pressure at each joint using at least one pressure sensor; a posture unit for calculating joint position or angle data using at least one flow measurement element; a controller generating a control signal for the valve; An interface that displays the robot's status and commands its movements; a 3D force estimation unit; The input of the 3D force estimation unit consists of a number of pressure sensors and a number of pose data; the output of the 3D force estimation unit consists of three outputs, one for each axis of force; The output of the 3D force estimation unit is applied to the end effector or gripper position; The output of the 3D force estimation unit is calculated from a neural network model obtained by supervised learning of a hydraulic drive system, a motor for electromechanically controlling the relief valve; The pressure sensor is disposed at an outlet of the relief valve; including the pressure sensor data at the input of the 3D force estimation unit; The pressure unit installed in one joint is obtained using two pressure sensors that measure the pressure state of the lower chamber and the rod chamber of the fluid actuator; The posture unit in one joint consists of two flow measurement elements, A hydraulic drive system characterized in that the attitude is calculated by subtracting a flow rate measured by one flow rate measuring element on the bottom surface of the fluid actuator and one flow rate measuring element on the rod side.

2. A hydraulic drive system equipped with a mobile manipulator consisting of a crawler and a manipulator, The manipulator includes at least one joint for rotating a first joint in an xz plane, at least one joint for rotating a second joint in another plane, and an end effector; A six-axis force torque sensor is installed on the end effector and is used only to obtain ground truth for learning data; the joint is actuated by at least one fluid actuator that moves a piston by supplying a fluid; a fluid driver for supplying and discharging fluid from a fluid reservoir to the fluid actuator; a pump for supplying fluid to the system from the fluid reservoir; a relief valve for adjusting the pressure of the fluid supplied from the pump to the system; one solenoid valve for controlling the flow of fluid from the system to the fluid actuator and another solenoid valve for controlling reverse flow; a pressure unit for measuring the fluid pressure at each joint using at least one pressure sensor; a posture unit for calculating joint position or angle data using at least one flow measurement element; a controller generating a control signal for the valve; An interface that displays the robot's status and commands its movements; a 3D force estimation unit; The input of the 3D force estimation unit consists of a number of pressure sensors and a number of pose data; the output of the 3D force estimation unit consists of three outputs, one for each axis of force; The output of the 3D force estimation unit is applied to the end effector or gripper position; The output of the 3D force estimation unit is calculated from a neural network model obtained by supervised learning of a hydraulic drive system, a motor for electromechanically controlling the relief valve; The pressure sensor is disposed at an outlet of the relief valve; including the pressure sensor data at the input of the 3D force estimation unit; The pressure unit installed in one joint is obtained using two pressure sensors that measure the pressure state of the lower chamber and the rod chamber of the fluid actuator; The posture unit in one joint consists of two flow measurement elements, 11. A hydraulic drive system, comprising: a 3D force estimation unit having an input from attitude data that is a number of flow measurement elements.

3. 3. The hydraulic drive system according to claim 2, the 3D force estimation unit is trained in three successive stages using training data from a data set according to a source of force applied to the end effector; 11. A hydraulic drive system, wherein the data set includes at least one of a force acting on a load, a disturbance force due to contact with the environment, and a disturbance force due to contact with the environment while holding a load.

4. 4. The hydraulic drive system according to claim 3, Further comprising a temperature sensor at the outlet of the relief valve; A hydraulic drive system, characterized in that the input of the 3D force estimation unit includes data from the temperature sensor.

5. 5. The hydraulic drive system according to claim 4, The front six-axis force torque sensor is calibrated in a state where it is attached to the end effector; Additional ground truth of calibration data is obtained when the manipulator is set to a predetermined pose; 11. A hydraulic drive system, comprising: a 3D force estimation unit that is trained in three successive stages using training data from one data set and calibration data for each of the sources of force applied by the end effector.

6. A hydraulic drive system equipped with a mobile manipulator consisting of a crawler and a manipulator, The manipulator includes at least one joint for rotating a first joint in an xz plane, at least one joint for rotating a second joint in another plane, and an end effector; A six-axis force torque sensor is installed on the end effector and is used only to obtain ground truth for learning data; the joint is actuated by at least one fluid actuator that moves a piston by supplying a fluid; a fluid driver for supplying and discharging fluid from a fluid reservoir to the fluid actuator; a pump for supplying fluid to the system from the fluid reservoir; a relief valve for adjusting the pressure of the fluid supplied from the pump to the system; one solenoid valve for controlling the flow of fluid from the system to the fluid actuator and another solenoid valve for controlling reverse flow; a pressure unit for measuring the fluid pressure at each joint using at least one pressure sensor; a posture unit for calculating joint position or angle data using at least one flow measurement element; a controller generating a control signal for the valve; An interface that displays the robot's status and commands its movements; a 3D force estimation unit; The input of the 3D force estimation unit consists of a number of pressure sensors and a number of pose data; the output of the 3D force estimation unit consists of three outputs, one for each axis of force; The output of the 3D force estimation unit is applied to the end effector or gripper position; The output of the 3D force estimation unit is calculated from a neural network model obtained by supervised learning of a hydraulic drive system, The manipulator includes at least one joint that rotates the first joint in an xz plane, at least one continuum joint that rotates the second joint in an xyz plane, and the end effector; The pressure unit measures the fluid pressure at the revolute joint by using at least one pressure sensor, and measures the fluid pressure at the continuum joint by using at least four pressure sensors; A hydraulic drive system, characterized in that the attitude unit calculates joint position or angle data at the rotary joint by using at least one flow measurement element, and calculates joint position or angle data at the continuum joint by using at least four flow measurement elements.

7. 7. The hydraulic drive system according to claim 6, a motor for electromechanically controlling the relief valve; The pressure sensor is installed near the outlet of the relief valve, an input of the 3D force estimation unit includes data from the pressure sensor; The pressure unit installed in the rotary joint is obtained using two pressure sensors that measure the pressure state of the lower chamber and the rod chamber of the fluid actuator; The pressure unit installed in the continuum joint is obtained using eight pressure sensors consisting of four sensors in the lower chamber of the fluid actuator and four sensors in the rod chamber; The attitude unit installed in the rotary joint is composed of two flow rate measuring elements that measure the flow of fluid in the lower chamber and the rod chamber of the fluid actuator; The attitude unit installed in the continuum joint is composed of eight flow rate measuring elements, which are four flow rate measuring elements in the lower chamber of the fluid actuator and four flow rate measuring elements in the rod chamber; 11. A hydraulic drive system, comprising: a 3D force estimation unit having an input from pose data that is a number of flow measurement elements.

8. 8. The hydraulic drive system according to claim 7, Further comprising a temperature sensor at the outlet of the relief valve; an input of the 3D force estimation unit includes data from the temperature sensor; The six-axis force torque sensor is calibrated while attached to the end effector; Once the manipulator is set to a predetermined pose, additional ground truth for training data is obtained; 11. A hydraulic drive system, comprising: a 3D force estimation unit that is trained in three successive stages using training data from one data set and calibration data for each of the sources of force applied by the end effector.

9. 7. The hydraulic drive system according to claim 6, a motor for electromechanically controlling the relief valve; The pressure sensor is installed near the outlet of the relief valve, an input of the 3D force estimation unit includes data from the pressure sensor; The pressure unit installed in the rotary joint is obtained using two pressure sensors that measure the pressure state of the lower chamber and the rod chamber of the fluid actuator; The pressure unit installed in the continuum joint is obtained using eight pressure sensors consisting of four sensors in the lower chamber of the fluid actuator and four sensors in the rod chamber; The attitude unit installed in the rotary joint is composed of two flow rate measuring elements that measure the flow of fluid in the lower chamber and the rod chamber of the fluid actuator; The attitude unit installed in the continuum joint is composed of eight flow rate measuring elements, which are four flow rate measuring elements in the lower chamber of the fluid actuator and four flow rate measuring elements in the rod chamber; 11. A hydraulic drive system, comprising: a 3D force estimation unit having an input from pose data that is a number of flow measurement elements.

10. 10. The hydraulic drive system of claim 9, Further comprising a temperature sensor at the outlet of the relief valve; an input of the 3D force estimation unit includes data from the temperature sensor; The six-axis force torque sensor is calibrated while attached to the end effector; Once the manipulator is set to a predetermined pose, additional ground truth for training data is obtained; 11. A hydraulic drive system, comprising: a 3D force estimation unit that is trained in three successive stages using training data from one data set and calibration data for each of the sources of force applied by the end effector.

Images