Search

JP-2026076276-A - Method for inferring force from a sensor device, method for training multiple networks, force inference module, and sensor device.

JP2026076276AJP 2026076276 AJP2026076276 AJP 2026076276AJP-2026076276-A

Abstract

[Challenge] Improve the configuration of sensors for robotics. [Solution] The present invention relates to a method for inferring force from a force-sensing sensor device. This method comprises reading out multiple pressure values and calculating a force map using a feedforward neural network. The present invention further relates to multiple methods for training multiple neural networks, a force inference module, and a sensor device. [Selection Diagram] Figure 9

Inventors

  • シュピアス・アダム
  • イ・ヒョサン
  • マルティウス・ゲオルグ
  • ソン・ファンボ
  • フィーネ・ヨナタン

Assignees

  • マックス－プランク－ゲゼルシヤフト・ツーア・フェルデルング・デア・ヴィッセンシャフテン・アインゲトラーゲナー・フェライン

Dates

Publication Date
20260511
Application Date
20260128

Claims (20)

  1. A method for inferring the force of a force-detecting sensor device (10), The sensor device (10) comprises a plurality of pressure sensors (400) and a follow-up layer (200) that covers the pressure sensors (400) and provides a measurement surface (210). The method for inferring the aforementioned force is, The steps include reading the pressure value (R) from the atmospheric pressure sensor (400), A step of calculating a force map (FM) on the measurement surface (210) based on the pressure value (R) using a feedforward neural network (FFNN), wherein the force map (FM) includes a plurality of force vectors (F), A method for inferring force, which includes the following features.
  2. The aforementioned feedforward neural network (FFNN) includes a transfer network (TN) and a reconstruction network (RN), The transfer network (TN) maps the pressure sensor (400) to a plurality of virtual sensors (400a) of the finite element model (10a) of the sensor device (10), The reconstruction network (RN) maps the virtual sensor (400a) of the finite element model (10a) to the force map (FM), Each of the virtual sensors (400a) comprises one or more virtual sensor points (410a) having a value (S). The method according to claim 1.
  3. Before force inference, the reconstruction network (RN) performs the following steps: - A step (T2_1) in which a plurality of simulations are performed in the finite element model (10a), Each of the simulations comprises simultaneously applying one or more simulated forces (605a) to the simulated measurement surface (210a) of the finite element model (10a), thereby calculating a simulated force map (FMa) on the simulated measurement surface (210a). The simulated force map (FMa) includes a plurality of simulated force vectors (F), Step (T2_1) involves calculating the value (S) of the corresponding virtual sensor point using the finite element model (10a), - A step (T2_2) of training the reconstructed network (RN) using the calculated simulated force map (FMa) and the corresponding calculated virtual sensor point values (S), The method according to claim 2, trained by
  4. The method according to claim 3, wherein the simulated force (605a) applied to the simulated measurement surface (210a) is generated based on each simulated indenter (600a) having a simulated indenter shape.
  5. The method according to claim 4, wherein the simulated indenter shape is selected from the group comprising at least a tip, a circle, a triangular cross-section, a square cross-section, a hemisphere, a cube, and a cylinder.
  6. The method according to any one of claims 3 to 5, wherein the reconstruction network (RN) is trained using a plurality of different simulated indenter shapes.
  7. The method according to any one of claims 3 to 6, wherein the reconstruction network (RN) is trained using a plurality of sized simulated indenters (600a).
  8. The method according to any one of claims 3 to 7, wherein the reconstruction network (RN) is trained by at least a portion of a simulation that includes the simultaneous application of simulated forces (605a) generated based on two or more simulated indenters (600a).
  9. The method according to any one of claims 3 to 8, wherein the reconstructed network (RN) is trained by at least a portion of a simulation that includes applying a simulated force (605a) generated based on only one simulated indenter (600a).
  10. The method according to any one of claims 3 to 9, wherein each simulated force vector (Fa) includes a normal force component ( FNa ), a first shear force component ( FS1a ), and a second shear force component ( FS2a ).
  11. In the simulated force vector (Fa), the first shear force component ( FS1a ) corresponds to the first shear force, and the second shear force component ( FS2a ) corresponds to the second shear force. The method according to claim 10, wherein the first shear force is perpendicular to the second shear force.
  12. The method according to any one of claims 3 to 11, wherein the reconstructed network (RN) is trained using a plurality of simulated forces (605a) having different shear force components.
  13. The method according to any one of claims 3 to 12, wherein the reconstruction network (RN) is trained using a plurality of simulated forces (605a) having different normal force components.
  14. Before force inference, the transfer network (TN) performs the following steps: - A step (T1_1) in which a plurality of force tests are performed on the sensor device (10), Step (T1_1) comprises each force test comprising applying a force with one indenter (600) at a position on the measurement surface (210) of the sensor device (10), simultaneously measuring the force (605) applied by the indenter (600), and simultaneously measuring the pressure value (R) with the barometric pressure sensor (400), - Step (T1_2) for each of the aforementioned force tests, in which a corresponding simulation is performed using the finite element model (10a), Each simulation includes applying a simulated force (605a) to the simulated measurement surface (210a) of the finite element model (10a), thereby calculating a simulated force map (FMa) on the simulated measurement surface (210a). The simulated force map (FMa) includes a plurality of simulated force vectors (Fa), Step (T1_2) is performed, wherein the simulated force (605a) is applied to a position on the simulated measuring surface (210a) that corresponds to the measured force (605) and the position on the measuring surface (210), The steps include: calculating the value (S) of the corresponding virtual sensor point using the finite element model (10a); - A step (T1_3) of training the transition network (TN) using the measured pressure value (R) and the corresponding calculated virtual sensor point value (S), The method according to any one of claims 2 to 13, trained in [the specified method].
  15. The method according to claim 14, wherein the force test for training the transfer network (TN) is performed using a plurality of indenters (600), each having its own indenter shape.
  16. The method according to claim 15, wherein the indenter shape is selected from the group comprising at least a tip, a circular shape, a triangular cross-section, a square cross-section, a hemisphere, a cube, and a cylinder.
  17. The method according to claim 15 or 16, wherein the simulation is performed by a simulated force (605a) based on a simulated indenter (600a) having a simulated indenter shape corresponding to the actual indenter shape used in the corresponding force test.
  18. The method according to any one of claims 14 to 17, wherein the transfer network (TN) is trained using a plurality of different indenter shapes.
  19. The method according to any one of claims 14 to 18, wherein the transfer network (TN) is trained using a plurality of indenters (600) of different sizes.
  20. The method according to any one of claims 14 to 19, wherein the transition network (TN) is trained by indenters (600) to which each shear force is applied for at least a portion of a force test training the transition network (TN).

Description

This invention relates to a method for inferring force from a sensor device, a method for training multiple networks, a force inference module, and a sensor device. When developing applications like robotics, detecting forces applied to parts of the robot, such as its hands, legs, or control devices, is crucial for enhancing its ability to move around and manipulate objects. Known implementations of sensor devices usable in robotic applications to obtain feedback on applied forces are extremely expensive and lack sufficient resolution. Such devices would be used to measure force. However, known sensor devices require a high density of sensors (or their arrangement) to achieve high spatial resolution. Figure 1 shows the arrangement of the sensors.Figure 2 shows the rigid core.Figure 3 shows a flexible circuit board.Figure 4 shows a rigid core with a flexible circuit board attached.Figure 5 shows an exploded view of the mold.Figure 6 shows the assembled state of the mold.Figure 7 shows an exploded view of the rigid core and mold covered with a flexible circuit board.Figure 8 shows the rigid core with the flexible circuit board attached, covering the molded pressure sensor.Figure 9 shows an overview of force inference.Figure 10 shows a finite element model.Figure 11 shows the arrangement of several different force particles.Figure 12 shows the setup for force testing.Figure 13 shows a flowchart for training the forwarding network.Figure 14 shows a flowchart for training the reconstructed network.Figure 15 shows a flowchart for training a feedforward neural network.Figure 16 shows a force map. Figure 1 shows a sensor device 10 according to one embodiment of the present invention. The sensor device 10 comprises a dome-shaped rigid core 100. The rigid core 100 is partially covered by a flexible circuit board 300 fixedly mounted to the rigid core 100. The flexible circuit board 300 is covered by a conforming layer 200. Multiple pressure sensors 400 are applied to the flexible circuit board 300. They protrude away from the rigid core 100. The conforming layer 200 provides a measuring surface 210 to which force can be applied. Because the conforming layer 200 is flexible and elastic, the force applied to the measuring surface 210 causes local deformation of the measuring surface 210, and the conforming layer 200 relays these forces to at least some of the pressure sensors 400. Therefore, the pressure sensors 400 can be used to evaluate force or applied force. The flexible circuit board 300 comprises multiple planar sections. These planar sections correspond to the structured planar sections on the rigid core 100, as shown in detail in Figure 2. The flexible circuit board 300 has a central portion 305 from which, in this embodiment, six arms extend. This central portion 305 can be considered a planar portion. All the arms are shown in Figure 3. In Figure 1, only three of these arms are visible, namely the first arm 310, the second arm 320, and the third arm 330, and are indicated by reference numerals. Each arm is divided into three planar sections; for example, the first arm 310 is divided into a first planar section 311, a second planar section 312, and a third planar section 313. The other arms are divided accordingly, and here, the planar sections 321, 322, 323, 331, 332, and 333 of the flexible circuit board 300 are visible in Figure 1. In the current embodiment, each planar section holds one pressure sensor 400. The central section 305 also holds one pressure sensor 400. Other configurations are possible; for example, the planar sections may have more than one pressure sensor 400, or they may not have any at all. Note that the pressure sensors 400 are spaced apart from each other on the flexible circuit board 300. However, much finer resolution regarding applied force can be achieved using the method described below. Figure 2 shows the rigid core 100 in sections. The rigid core 100 consists of a total of six surface areas, of which the first surface area 110, the second surface area 120, and the third surface area 130 are visible and shown in Figure 3. Each surface area 110, 120, and 130 is divided into three planar sections. For example, the first surface area 110 is divided into the first planar section 111, the second planar section 112, and the third planar section 113. The other surface areas are divided accordingly, with planar sections 121, 122, 123, 131, 132, and 133 visible in Figure 1. At the top of the rigid core 100, the central section 105 connects the multiple surface areas. The planar portion of the rigid core 100 defines the planar portion of the flexible circuit board 300. Specifically, the planar portions have different orientations, and the flexible circuit board 300 adapts to each orientation of the planar portion. Figure 2 clearly shows that the rigid core 100 is dome-shaped, and the rigid core can be used, for example, in the fingertips of a robot. Figure 3 separately shows a flexible circ