Back to Bigger Picture

BIGGER PICTURE Accessibility & independence

A prosthetic knee learned its wearer's walk in about ten minutes

A powered knee can supply movement, but its settings must fit the person wearing it. Researchers explored whether the prosthesis could learn that fit while its user was already walking.

Three successive views of a participant walking on a treadmill with a powered prosthetic knee.Original photograph
Treadmill testing of a powered prosthetic knee in the research on using reinforcement learning to tune the device. The montage shows successive steps.Helen HuangEditorial reuse with attribution
Adjustable settings
12
Average tuning time
About 10 minutes
Initial study
2 participants on a treadmill

How the idea works

  1. Measure the step

    Knee sensors record movement through the gait cycle.

  2. Learn the adjustment

    Reinforcement learning compares knee motion with its target.

  3. Retune while walking

    The controller updates stiffness and other parameters across the step.

The person changes along with the machine

A robotic knee needs different stiffness and movement settings during different parts of a step. Adjusting them is difficult because the person also changes how they walk in response to the prosthesis. A fixed mechanical model cannot fully describe that interaction.

Sources: [1][2]

Each walk becomes feedback

The team used reinforcement learning to adjust twelve control parameters. Sensors recorded the knee's movement; the algorithm compared it with a target walking pattern, learned which adjustments helped, and tried improved settings while the participant continued walking.

Sources: [1][2]

A small experiment with a practical result

In one participant with an amputation and one without, the system reached its target knee motion after an average of roughly 300 gait cycles, or ten minutes. The 2019 study demonstrated automatic tuning on a treadmill at a fixed speed.

Sources: [1]

The next challenge is what the wearer wants

The target was a reference knee movement, not a complete measure of comfort, balance or personal preference. Stairs and everyday walking were outside this experiment. Its contribution is a way to reduce repeated manual adjustment while allowing a person and an assistive machine to adapt together.

Sources: [2]

The original sources

  1. IEEE Transactions on CyberneticsOnline Reinforcement Learning Control for the Personalization of a Robotic Knee Prosthesis
  2. North Carolina State UniversityReinforcement Learning Expedites ‘Tuning’ of Robotic Prosthetics

What to explore next

A robotic boot learns from forty-four steps at a timeListening for a volcano’s change of rhythm
Explore all 99 applications