Learning-based tensegrity robot joint

1 . A tensegrity joint comprising:

a first inner structure having a first frame and a first strut extending orthogonally from the first frame;

a second inner structure having a second frame and a second strut extending orthogonally from the second frame;

at least three actuation tendons, each being fixed at a respective hole at the first frame and being guided and allowed to slide through a respective hole at the second frame;

a central tendon connecting an apex of the first strut and an apex of the second strut such that the central tendon is in balance with the pulling forces exerted by the at least three actuation tendons;

an outer elastic protection sheath configured to wrap around a cylindrical surface between the first and second inner structures; and

a fiber-Bragg-gratings sensor wounded helically around and embedded inside the outer elastic protection sheath and configured for sensing configurations of the tensegrity joint in real time for closed-loop feedback control.

2 . The tensegrity joint of claim 1 , wherein the outer elastic protection sheath is prestressed to pull the first and second inner structures to each other to provide resistance to torsional force and restore the tensegrity joint to an original or natural equilibrium.

3 . The tensegrity joint of claim 1 , wherein the first strut and the second strut are positioned relative to each other at a level to allow the tensegrity joint to bend at a bending angle.

4 . The tensegrity joint of claim 1 , wherein each of the actuation tendons and central tendon is made of braided metal wires.

5 . The tensegrity joint of claim 1 , wherein the first and second frames are circular.

6 . The tensegrity joint of claim 1 , wherein the first and second struts are “L”-shaped.

7 . The learning-based tensegrity robot joint of claim 1 , wherein the outer elastic protection sheath is made of silicone.

8 . A learning-based tensegrity robot joint, comprising:

the tensegrity joint of claim 1 ; and

a learning-based controller configured for controlling the tensegrity joint.

9 . The learning-based tensegrity robot joint of claim 8 , wherein the outer elastic protection sheath is prestressed to pull the first and second inner structures to each other to provide resistance to torsional force and restore the tensegrity joint to an original or natural equilibrium.

10 . The learning-based tensegrity robot joint of claim 8 , wherein the first strut and the second strut are positioned relative to each other at a level to allow the tensegrity joint to bend at a bending angle.

11 . The learning-based tensegrity robot joint of claim 8 , wherein each of the actuation tendons and central tendon is made of braided metal wires.

12 . The learning-based tensegrity robot joint of claim 8 , wherein the first and second frames are circular.

13 . The learning-based tensegrity robot joint of claim 8 , wherein the first and second struts are “L”-shaped.

14 . The learning-based tensegrity robot joint of claim 8 , wherein the outer elastic protection sheath is made of silicone.

15 . The learning-based tensegrity robot joint of claim 8 , further comprising at least three motors for pulling the at least three actuation tendons respectively.

16 . The learning-based tensegrity robot joint of claim 8 , wherein the learning-based controller is trained on basis of a training dataset prepared by a sampling technique including:

generating, based on a geometric model of the tensegrity joint, an actuation sequence containing initial tendon configurations; and

executing all the initial tendon configurations contained in the generated actuation sequence and correcting, during actual joint movement, each of the initial tendon configurations in a torque-control manner by multiple trials of releasing and pulling the tendon individually.

17 . The learning-based tensegrity robot joint of claim 16 , wherein the learning-based controller is further trained on basis of sensing feedback provided by the fiber-Bragg-gratings sensor.

18 . A method for training the learning-based tensegrity robot joint of claim 8 , the method comprising:

training the learning-based controller of the learning-based tensegrity robot joint on basis of a first training dataset prepared by a sampling technique;

installing an end-effector on the learning-based tensegrity robot joint;

commanding the learning-based controller to move the end-effector to move along a spiral path for a plurality of pre-set distances within a workspace;

tracking a tip position of the end-effector by a tracker and detecting a strain change of the outer elastic protection sheath by the fiber-Bragg-gratings sensor after the end-effector is moved with each of the pre-set distances; and

collecting the tracked tip positions and detected strain changes to form a second training dataset to train the learning-based controller to operate as a close-loop controller.

19 . The method according to claim 18 , wherein the geometric model is deduced by assuming:

there is no axial rotation of the central tendon in the absence of external disturbance; and

the central tendon is perpendicular to the first frame.

20 . The method according to claim 16 , wherein the sampling technique includes:

generating, based on a geometric model of the tensegrity joint, an actuation sequence containing initial tendon configurations; and

executing all the initial tendon configurations contained in the generated actuation sequence and correcting, during actual joint movement, each of the initial tendon configurations in a torque-control manner by multiple trials of releasing and pulling the tendon individually.