Step S120 also specifically includes the following steps of S121 to S123.

At S121, the tactile data is filtered by a Kalman filter, so that the tactile flow data of the robot along the direction of the assembly action is obtained by means of an optical flow algorithm (Farneback algorithm). As shown in FIG. 1, in the embodiment described above, after the acquired tactile information is filtered by the Kalman filter, and, analogous to the optical flow, the tactile flow in the x-axis and y-axis directions of the clamping effector in each action step is calculated by the Farneback algorithm (with a dimension of 25 (5×5 detection points)×2 (both directions of the x-axis and y-axis)×10 (due to every tactile flow information being calculated for every tactile information of every two consecutive waypoints). In addition, the tactile flow data can be reduced by causal convolution with a dimension of 25 tactile detection points×2 tactile detection points to obtain processed tactile data.

At S122, the captured RGB image data is converted to a grayscale image, and for the last waypoint in each assembly action step of the clamping effector, the RGB image data is compressed to an RGB representation vector by an RGB image encoder, and the RGB representation vector is clipped by an RGB image decoder to a processed grayscale image with 128 pixels×128 pixels.

At S123, the captured depth image data is converted to a grayscale image, and for the last waypoint in each assembly action step of the clamping effector, the depth image data is compressed to a depth representation vector by a depth image encoder, and the depth representation vector is clipped by a depth image decoder to a processed depth image with 128 pixels×128 pixels.

As shown in FIG. 1, in steps S122 and S123, in the embodiment described above, the RGB image and the depth image of the plug and socket, which are output by the camera around the assembly area, are clipped to a size of 128 pixels×128 pixels, and the RGB image is converted to a grayscale image, so as to reduce the size and processing time of the data.

At S124, a variational auto encoder is provided, to which the processed grayscale image and the processed depth image are input, for training multiple times respectively, and outputting an RGB representation vector with a dimension of 16×1 and a depth representation vector with a dimension of 16×1 respectively. In the above embodiment, both the grayscale image and the depth image of the plug and socket assembly, are used for training by a variational autocoder (VAE), thereby the images should be encoded and compressed into a 16×1 “representation vector” ensuring that the most of information of the original image is covered.

Embodiments of Steps S130 and S140

The dimensions of the tactile data representation vector and the assembly force data of the robot are descended by causal convolution. The descended data is input to the tensor fusion network model for determining whether the clamping effector reaches an assembly damping point by a multi-layer perceptron including a damping point predictor.

More specifically, for each action step, the tactile flow is descended to a size (i.e., dimension 25×2) by means of a causal convolutional network, and then fused with the representation vector of the two images (i.e., two 16×1 vector) in step S120 by means of a tensor fusion network model (TFN) (i.e., Kronecker's product is performed by adding an element “1” to the end of each modal vector, and the result obtained will contain both the single information and the arbitrary combination of the information of these three modalities), and then the resulting three-dimensional tensor is spread out and inputted into the damper-point predictor including a multi-layer perceptron, which is used for predicting whether or not a plug arrives at the snap damper-point in a socket.

Combining the above embodiments, the plugs and exits extract the tactile data of the 5 action detection points×5 action detection points of the last waypoint in each action step, which are descended using Gaussian convolution (i.e., with a dimension of 9×1), fused with the representation vector of the two images by direct splicing (i.e., with a dimension of 16+16+9), and the resulting representation vector (i.e., with a dimension of 41×1) are reintroduced into the reinforcement learning network and generate the action values for the clamping effector (the five action directions x, y, R, P, Y).

Embodiments of Step S150

In some embodiments, the torque of the 6 movable joints of the robot as shown in F. 4 is solved by the following Jacobi matrix:

τ
q

=

K
p

⁢

J

-
1

⁢

S
¯

⁢
J
⁢
Δ
⁢
q

+

K
v

⁢

J

-
1

⁢

S
¯

⁢
Δ
⁢

q
˙

+

J
T

(

K
Fp

⁢
S
⁢
Δ
⁢
F

+

K
Fi

⁢
S
⁢

∫

0

t

Δ
⁢
F
⁢
d
⁢
t

+

F
d

)

Wherein, K_p, K_v, K_Fpand K_Piare the control parameters for proportional differentiation (PD) and proportional integration (PI), respectively. Sis the discriminant matrixS=diagS=diag(1, 1, 0, 1, 1, 1), that is, the bit control dimension corresponds to a value of 1 and the force control dimension corresponds to a value of 0.

The robot controller implements position control by means of a PD control algorithm; the robot controller implements force control by means of a PI control algorithm.

With reference to the schematic diagrams of the curves of the reinforcement learning results according to the method of the disclosure shown in FIGS. 5