Selected researchRobotics

Coordinated motion in tendon-driven arms

Tendon-driven robotics ยท Ongoing research

Shiqi Ren

One physical arm section with a white cylindrical frame, guide disks at both ends and wiring through the central opening.
An assembled arm section. The guide disks carry tendons along the arm, while the motors that pull them sit in a separate drive box.

How can one section move on its own?

The arm has six sections connected by universal joints. Each joint can pitch and yaw, giving the arm twelve rotational degrees of freedom. Eighteen tendons run from the drive box, with three terminating at each section.

Side view of the straight six-section arm assembly, showing repeated structural segments, guide disks and connecting joints.
Mechanical design of the full arm in its straight configuration. A tendon that ends near the tip passes through several upstream guide disks.

Bending an upstream joint changes the length of the route to a downstream section. That coupling gives us a concrete control problem. Can we choose a set of tendon length or tension changes that bends one section while limiting motion in its neighbors? I am studying how the routing geometry constrains those choices.

When one tendon loses effectiveness

In a separate single-section simulation, one tendon suddenly transmits only part of the force requested by the controller. The controller observes the motion but does not know which tendon weakened. One strategy uses the mismatch between predicted and observed motion to identify the affected tendon and reassign tension.

Episodes that met the tracking and recovery criteria
ControllerLocal modelNonlinear model
Nominal tension allocation4 / 241 / 4
Identify and reassign tension24 / 244 / 4
Learned residual, one observation0 / 240 / 4
Learned residual, eight observations0 / 240 / 4

The local-model set included 4 healthy and 20 fault episodes. The nonlinear check used 1 healthy and 3 fault episodes. Both used noiseless simulated observations. The two learned policies each trained for 50,000 steps with one seed and had a narrower tension adjustment range than the allocator. The next comparison will give the controllers the same range.