Sensing long cane · Prototype · 2026
Notices the kerb, the step and the branch before you reach them, and tells your hand.
It adds what the tip can't reach: the drop a metre ahead, and the branch at head height.
An 8 × 8 depth sensor finds ground that is further away than it should be.
Simulated: 1 m+ ahead of a 100 mm kerb
Ground that rises too early is a step or an obstacle, each with its own rhythm.
Confirmed on 2 of 3 frames
Ultrasound covers chest and head height, where the tip never reaches.
Range 0.8 – 2.5 m
Ground cues under the fingers, overhead under the thumb. Faster means closer.
Motors 2 · rhythms 4
A dropped cane is let go before it lands; a falling person holds on. Then it alarms.
Grip read every 10 ms
One 18650 cell and USB-C. It sleeps when you put it down.
Estimated ≈ 10 h
A few degrees of tilt make flat ground look a kerb deep, so NORDIC never thresholds distance. It predicts where the ground should be in each of 64 zones and watches for it to disagree.
Gyro and accelerometer give the pod's tilt and sweep.
Each zone gets the range flat ground would return.
64 real ranges, 15 times a second.
Further is a drop; closer is a step or an object.
Two of three frames agree, then the motor plays.
4 : 1
At 1.5 m, one degree of tilt error looks like a 26 mm step. A kerb is 100 mm. That margin is why attitude must be good to 1°.
Direct sun blinds the depth sensor, so the cane gives a warning cue instead of going quiet. In shade, warnings come under a metre ahead. Very deep drops and corridor walls are still open problems. Ultrasound works in any light.
Five systems, each designed, simulated and laid out on our own two boards.
Where you feel it says where the hazard is. The rhythm says what it is; the repeat rate, how close.
Three windings under the grip time an oscillator at about 80 kHz. A hand pulls the frequency down; a driven guard hides the electronics inside.
The accelerometer knows down but the sweep throws it off; the gyro is smooth but drifts. A complementary filter takes the best of each.
A protected cell, a buck-boost regulator and a hardware latch that turns the processor fully off. Asleep, only the motion sensor listens.
The handle board carries the processor, power and motors; the pod board carries the sensors. One flat cable joins them.
Animated from our design values. Anything not yet measured on the bench is illustrative.
Our design targets, and where each one stands today.
| Measure | Target | Where it stands |
|---|---|---|
| Drop-off warning | ≥ 1.0 m ahead of the leading foot | Simulated ≥ 1.05 m indoors; under 1.0 m in shade |
| Detection rate | ≥ 90 % of approaches | Simulated 148 of 148 indoors |
| Smallest ground change | 100 mm | By design Threshold from 50 mm |
| Overhead range | 0.8 – 2.0 m | Simulated 1.70 m bar warned 1.5 m ahead |
| False alarms | ≤ 2 per 100 m | Open None indoors; 1 – 5 in shade; walls too many |
| Depth frames | 15 Hz | By design 15 Hz, 8 × 8 zones |
| Detection to cue | ≤ 300 ms | Estimated ≈ 200 – 250 ms |
| Battery life | ≥ 8 h | Estimated ≈ 10 h |
| Standby power | < 4 mW | Estimated ≈ 0.6 – 0.9 mW |
| Charge to 90 % | ≤ 3 h | At risk ≈ 3.2 – 3.6 h at 750 mA |
| Calibration | ≤ 60 s | To test Guided, three steps |
| Fall detection | ≥ 90 % of staged falls | To test Grip timing plus motion |
| Alarm loudness | ≥ 85 dB at 1 m | Below target ≈ 68 – 69 dB, datasheet |
Drawn from our own design files. Select a marker to zoom in.
A · Circuit boards
Four layers, designed in KiCad, made by JLCPCB. Rev H, 6 Oct.
5 A fuse · charger · buck-boost to 3.3 V
The cell meets a fuse first, then the regulator over a short, wide trace.
2 × DRI0056 · own I²C bus each
The two haptic driver boards hang under the board on these mounts.
ESP32-P4 · power latch · cell protection
The module plugs in on top; the latch and battery protection sit beneath it. Off means fully off.
Schmitt oscillator · guard follower
No ground pour under the sense node on any layer, and its own clearance rule.
20-way FFC · 0.5 m · buffered
Every signal passes a buffer that lets go when the processor is off.
Right-angle · 2 mm tall · 20 ways
Lets the cable run straight back along the shaft.
Load switch · 1.9 ms soft start
Depth sensor and sonar are cut off when idle; the motion sensor stays on to wake the cane.
One SPI bus · two sensors
Keep the processor's signals out of an unpowered sensor.
Depth · sonar · 10 µF + 100 nF each
Every sensor wire has a series resistor and local decoupling.
B · Grip sensor
A hand adds a few picofarads and the oscillator slows.
≈ 80 – 85 kHz hands off, expected
There is no timing capacitor: the sense winding is it, so a hand slows the oscillator directly.
MCP6022 follower · ≤ 150 pF coupling
The guard winding follows the sense voltage, so the electronics inside the grip are invisible.
≈ 2 × signal on battery, predicted
Gives a floating, battery-powered cane a return path through the palm.
74HC14 + 750 kΩ · 10 ms gate
The processor counts edges for 10 ms. Held is more than 6 % down.
Values as fitted on 8 Oct. Hand figures are design estimates and simulation.
C · Mechanical
Cut from the 15 Oct CAD. Hatched faces are printed PETG.
Clamshell · Ø 8 mm carbon shaft
Two printed halves close on the shaft and carry the buttons and the guarded rocker.
Windings on the shell · guard on the core
The sense and ground windings are wound straight onto the printed shell.
2 × LRA · 10 × 10 × 3.5 mm
One on the thumb side, one under the fingers, so each cue lands where it means something.
18650 · 197 mm board, end to end
The weight sits in your hand, not out along the shaft.
4 × M5 · no glue
Base and clamp cap close on the shaft, so the pod can be set for each cane.
Motion sensor + depth sensor, no joint
Nothing flexes between them, so the tilt measured is the tilt the depth sensor has.
45° field · 26° – 42° below horizontal
Set to the shaft, not the horizon, so the ground stays in view for any pitch from 27° to 57°.
25° – 9° above horizontal
Looks just above horizontal, at chest and head height.
“In use” is drawn at 44.5°, a mid grip. Beam widths are cut short.