Adaptive Prosthetic Ankle Control for Terrain Changes

ISEF Category: Biomedical Engineering

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A prosthetic ankle can feel fine on a flat hallway and fail on a curb. That jump matters because tiny changes in terrain can throw off balance, comfort, and energy use. Your project asks whether a smart controller can notice the ground and adjust fast enough to help.

What Is It?

This project studies a prosthetic ankle that changes how it moves based on the ground under it. Think of it like cruise control for walking, but one that also reads the road. A reinforcement-learning controller learns which ankle actions work best in different terrain classes, such as flat ground, stairs, ramps, or uneven surfaces.

The terrain signal comes from a low-cost IMU, which measures motion and orientation, plus a foot camera, which gives visual clues about the surface. The controller uses that input to pick an ankle policy in real time. In plain terms, it tries to answer, “What should the ankle do right now?” before the step is finished.

You then compare the simulated performance to published gait data from passive ankles. Passive ankles store and return energy, but they do not adapt much. Your job is to see whether a terrain-aware active design can improve stability, smoothness, or symmetry when the walking surface changes.

Why This Is a Good Topic

This is a strong science fair topic because you can test it with clear inputs, clear outputs, and published benchmarks. You can change the terrain class, the sensor setup, or the controller rule, then measure how those changes affect gait quality. The project connects to real prosthetics, fall prevention, and wearable robotics. You can also learn simulation, sensor fusion, machine learning, and data analysis in one project.

Research Questions

• How does terrain-class accuracy from a low-cost IMU and foot camera affect prosthetic ankle control performance?
• What is the effect of using only IMU data versus IMU plus camera data on terrain classification accuracy?
• Does a reinforcement-learning ankle policy improve gait symmetry compared with a passive-ankle baseline on mixed terrain?
• To what extent does controller adaptation time change when the terrain switches from flat ground to ramps or stairs?
• Which terrain features are most predictive of ankle policy choice in simulation?
• How does sensor noise in the IMU or camera change the controller's stability metrics?
• What is the effect of training on one set of terrains and testing on a new terrain on policy generalization?

Basic Materials

• Computer with enough memory to run MuJoCo or a similar physics simulator.
• MuJoCo software license or academic access.
• Python installed with data analysis libraries.
• Low-cost IMU sensor module.
• Small foot-mounted camera or webcam for prototype testing.
• 3D-printed or cardboard foot model for sensor placement tests.
• Published passive-ankle gait data from journal articles or open datasets.
• Spreadsheet or notebook for tracking experiments and results.
• External storage or cloud backup for simulation logs.

Advanced Materials

• University workstation with a GPU for training reinforcement-learning models.
• Motion capture system or instrumented treadmill for validation.
• Research-grade IMU sensors.
• High-speed foot camera or wearable vision module.
• Prosthetic ankle test bench or robotic ankle hardware.
• Force plate or pressure insoles for ground-truth gait metrics.
• Biomechanics dataset with joint angle, torque, and ground reaction force data.
• MATLAB or Python environment for signal processing and statistics.
• Access to IRB guidance if any human walking data is collected.

Software & Tools

• Python: Runs simulation control, data processing, and reinforcement-learning experiments.
• MuJoCo: Simulates gait, ankle dynamics, and terrain interactions.
• ImageJ: Measures and compares visual features if you test camera-based terrain sensing.
• Jupyter Notebook: Helps you organize results, plots, and model comparisons.
• R: Supports statistical tests and graphing for gait and classification results.

Experiment Steps

1. Define the gait outcome you care about, such as stability, symmetry, or energy cost.
2. Choose the terrain classes your controller must recognize and explain why those classes matter.
3. Build a baseline model that uses a passive-ankle rule or a fixed active policy.
4. Add terrain detection and decide how the detected class changes the ankle action.
5. Plan a comparison framework that uses the same walking conditions for every controller.
6. Select evaluation metrics that let you compare your simulation to published gait data.

Common Pitfalls

• Training the controller on terrain labels that are too easy, which makes the model look good in simulation but fail on new surfaces.
• Letting the camera dominate the IMU signal, which can hide the effect of motion sensing and weaken generalization.
• Comparing your active policy to a passive baseline that uses a different speed or step pattern, which makes the results unfair.
• Using published gait data without matching the same terrain class, which breaks the benchmark.
• Ignoring sensor noise and misclassification, which makes the controller look more stable than it really is.

What Makes This Competitive

A competitive version of this project would do more than train a model and report accuracy. You would test how well the controller handles noisy sensors, unseen terrain, and fair baseline comparisons. Strong entries also use clean metrics, like step-to-step stability, symmetry, and error under terrain switches. If you can connect simulation results to published biomechanics data in a careful way, your project will feel much closer to real research.

Project Variations

• Use only IMU data and compare it with IMU plus camera input to test whether vision actually helps terrain detection.
• Swap the terrain classifier for a direct end-to-end policy and compare adaptation speed against a two-stage sensing model.
• Benchmark your active ankle against different passive-ankle datasets, such as level walking, ramps, or stairs, to see where adaptation matters most.

Learn More

• MuJoCo Documentation: Official simulator docs that explain model building, physics settings, and control loops. Find it by searching for the MuJoCo documentation site.
• NIH PubMed: Search review articles on prosthetic ankle control, gait adaptation, and reinforcement learning in rehabilitation robotics.
• PubMed Central: Find full-text papers on prosthetic biomechanics and wearable sensing when abstracts are not enough.
• NASA OpenMDAO or Python tutorials: Use open Python resources for workflow design, logging, and optimization-style thinking. Find them through official documentation pages.
• MIT OpenCourseWare: Search biomechanics, robotics, and control courses for free lecture notes and assignments.
• IEEE Xplore abstract pages: Read abstracts and, when available, open-access papers on prosthetic control and terrain classification.