Battery-Free Smart Shoes Run AI Via Walking Power

Smart Shoes Powered by Footsteps Can Tell Walking, Running, and Stair Climbing Apart

A smartwatch that turns off mid-workout is a familiar frustration. A prototype pair of sneakers sidesteps the problem with no battery and no charging cable, yet it can tell walking from running from stair climbing, count steps, and estimate calories burned. Engineers led by a Rutgers University team report in Science Advances that the shoe’s motion sensor and artificial intelligence (AI) software run on electricity generated by the wearer’s own footsteps, drawing an average of just 86 microwatts of power.

Comparable wearable AI systems typically run in the 10 to 100 milliwatt range, the authors report, making this design roughly a hundred times leaner. A milliwatt is one-thousandth of a watt, and a microwatt is one-millionth.

Each footstep squeezes and releases an energy-harvesting layer tucked into the shoe’s air cushion, and the team built the electronics to run on that pulsed current. Operation continues as long as the wearer keeps moving, and the paper does not show uninterrupted operation through long motionless stretches such as sleep. Even so, a battery-free design targets a real problem, since a dead battery interrupts monitoring in devices meant to track walking over long periods.

Battery-Free Smart Shoes Can Start From Zero Charge in as Little as 5.95 Seconds of Lab Tapping

Four connected parts make up the system. An energy harvester, a layered material that produces a small electrical charge each time its layers press together and spring apart, sits in the sole’s air cushion, measures about 3.1 by 2.3 inches, and weighs under 2 ounces. A power management circuit smooths that jumpy trickle into steady power. An accelerometer, a tiny chip that senses motion in three directions, gathers movement data. A compact AI program on the shoe’s built-in computer chip interprets it.

Starting from nothing was its own puzzle. A device with a fully drained battery cannot turn itself back on, and control circuits normally need some power just to wake up. The team added a cold-start module that captures the first small bursts of energy from footsteps, uses them to wake the circuit, then switches to a more efficient charging route. In a laboratory test where an operator tapped the harvester as fast as possible, the power management circuit started from zero stored energy in as little as 5.95 seconds.

Counting the power management circuitry, the full system drew about 108 microwatts, still well below what the harvester produced in lab tapping tests.

Even Slow Walking Can Power the Battery-Free Smart Shoes’ Entire System

Four healthy volunteers, three men and one woman between ages 23 and 26, wore the shoes. They stood roughly 5 feet 5 inches to 5 feet 11 inches tall and weighed about 159 to 205 pounds. Each used a treadmill set to 1.3 mph for slow walking, 2.0 mph for fast walking, and 3.0 mph for running, staying at each pace for several minutes. A stair climber set to 60 steps per minute covered the stair-climbing trials.

After trimming the start and end of each trial, the team sliced the recordings into 537 windows of 15 seconds each. About 80% trained the AI, while the remaining 20%, held back until later, tested it. The accelerometer’s readings were also checked against an iPhone’s built-in motion sensor during side-by-side shaking tests.

In intended use, raw movement data does not need to leave the shoe. The chip processes each window itself and shows finished results, including step count, activity type, and estimated calories, on a small low-power screen mounted on the shoe. The authors chose a screen over wireless streaming because radio communication would demand far more energy, and future versions could send compact summaries wirelessly. Step counts from the left and right shoes matched closely and agreed with hand counts, according to the authors. Even slow walking can sustainably power the entire system, the paper reports.

A Slimmed-Down AI Model Runs 15 Times Faster and Reaches 95.4% Accuracy on Test Data

Fitting AI onto a chip that sips power took trimming. An early version of the activity classifier was slightly more accurate but needed more memory than the chip could hold and took 6.1 seconds to process each 15-second window of data.

A technique that ranks which pieces of motion data matter most for telling activities apart pointed to just three measurements: how much the foot’s acceleration varies along each of three directions. Dropping everything else also let the researchers skip a math-heavy processing step. That slimmed-down version classified the four activities with 95.4% accuracy on the study’s test data. Processing time fell to 0.4 seconds per window, and electrical current dropped more than sixfold.

Calorie estimates come from a standard formula that multiplies body weight, time spent, and a preset intensity value for each activity. The program assumed 75 kilograms (about 165 pounds) when a wearer’s weight was unavailable, so the numbers are estimates, not direct measurements.

Gait, or the way a person walks, serves as a warning sign for conditions including cystic fibrosis, Parkinson’s disease, spinal cord injury, and rare bone diseases, the authors note. They see the platform eventually tracking outcomes such as fall risk or recovery, though that would require retraining and validating the AI with larger, more varied groups, including people with health conditions.

For now, the shoe shows that AI activity tracking can run on walking power alone. Whether it holds up in patients remains untested.

---