Simulating a cable-driven parallel robot that assists only as needed
In motor rehabilitation, a robot that does the whole movement for the patient is of little use: recovery depends on active participation. This work builds a simulation environment in which an eight-cable parallel robot assists the arm of a virtual patient and adjusts its help from a single signal, tracking error, without knowing the patient's motor capacity. Across three simulated profiles, robot participation went from 20.4% with a healthy subject to 100% with a severely impaired patient, without changing any controller parameter.
Context
Cable-driven parallel robots move an end effector by pulling on it with several cables from a fixed frame. They have little moving mass, a large workspace and a structure that does not surround the patient with rigid links. Their fundamental constraint is that a cable can only pull, so not every force is achievable at every position.
Most simulation studies represent the patient in a very simplified way, either fully passive or fully active. That simplification prevents evaluating exactly what matters: how effort is shared between the person and the machine.
Components
Stages of the work, from motion capture to cable geometry optimization
Biomechanical model. A musculoskeletal upper-limb model in MuJoCo, compared against motion capture of elbow flexion, pronation-supination and wrist flexion-extension.
Synthetic patient. Motor impairment is represented by a position-dependent functional coefficient: the available voluntary torque decreases progressively as the movement exceeds the person's functional range. Three profiles were defined, with limits of 90° (severe), 120° (moderate) and 150° (healthy reference).
Reference trajectory and synthetic impairment profiles, together with motion capture data
Robot. Eight cables on a cubic frame. Its tracking was characterized in isolation over five trajectories (line, circle, lemniscate, helix and spherical spiral), with Cartesian errors between 0.44 and 1.92 mm.
Lemniscate trajectory: tracking error and tensions of the eight cables
Controller. A three-layer architecture: an adaptive impedance controller estimates the required assistance from the error, that assistance is converted into a force on the end effector, and the force is distributed among the cables through the pseudoinverse of the structure matrix, with iterative redistribution to keep all tensions positive.
Results
Controller response for the moderately impaired patient in elbow flexion
In all three profiles the angular error stayed below 1.1° and tensions below 30 N. The central result is that assistance emerges from the interaction: the controller receives no information about the patient's capacity and still compensates exactly what the patient fails to do.
Torque sharing between patient and robot for the severe profile
The same scheme worked without retuning for wrist flexion-extension, radial-ulnar deviation and shoulder rotation. Pronation-supination turned out to be geometrically infeasible with cables, and the work recommends direct actuation for that degree of freedom.
The anchor geometry was optimized with NSGA-II. Coverage of the workspace in which the required forces are achievable rose from 77.2% to 88.2%, and going from six to eight cables brought diminishing improvements.
Pareto fronts for 4 to 8 cables: workspace coverage against peak tension
What is missing
Everything happens in simulation. The physical prototype that was built demonstrates mechanical feasibility but runs in open loop.
Physical proof-of-concept prototype
The patient model is deterministic and does not include fatigue, velocity-dependent spasticity, co-contraction or session-to-session variability. The author is explicit about this: it is a reproducible test bench for controllers, not a physiological model. Cable elasticity, friction and sensor noise were not modeled either.
How it fits in Robiolab
The work combines two of the lab's interests: robots actuated by cables and tendons, such as the tendon-driven continuum robot, and assistance of human movement. Its contribution to the group is a platform on which a control strategy can be evaluated against different levels of impairment before building hardware or involving patients.