DRESF 2026–27 · Engineering

A robot that watches over the people you love.

SafePath patrols a home on its own, recognizes when someone has fallen using AI pose estimation, raises a camera mast to verify — then alerts family in seconds. Designed, wired, and built from the cells of the battery up.

▶ Run the live demo Engineering details
Photo — full robot, front 3/4 view
camera mast · 10 ft reach custom 24V pack · 14 cells
0onboard computers
0 ftcamera mast reach
0 Vhand-built battery pack
< 0 sfall-to-alert target
Photo — the build in progress
01 Why SafePath exists

Built because of my grandmother.

Every year, over 3 million older adults are treated in emergency rooms for fall injuries — and when someone lives alone, the most dangerous part isn't the fall. It's the hours nobody knows it happened.

SafePath started as a question about my own grandmother: if she fell and couldn't reach a phone, how long would it take for anyone to find out? Wearable pendants only work if you wear them and can press them. I wanted something that watches out for you even when you can't call for help yourself.

A safety companion that doesn't need to be worn, charged nightly, or remembered — it just patrols, watches, and acts.
02 How it works

Detect. Verify. Extend. Alert.

A real emergency response has to avoid false alarms without missing real falls. SafePath runs a four-stage pipeline — each stage has to agree before help is called. Click a stage, or watch it cycle.

STAGE 1

Detect

An onboard camera feeds a YOLOv8 pose-estimation model running on the robot itself — no cloud. The AI tracks 17 body keypoints and flags postures consistent with a fall.

STAGE 2

Verify

Lying on the couch isn't an emergency. SafePath checks how the person got there — sudden drop vs. slow sit — and how long they stay motionless before treating it as real.

STAGE 3

Extend

A motorized 4-stage mast raises the camera up to 10 feet for a clear vantage over furniture, confirming the person's state and capturing a snapshot for responders.

STAGE 4

Alert

The touchscreen asks "Are you okay?" with a big response button. No response in time → family gets an app alert with location and snapshot.

03 System architecture — live demo

Three computers, one nervous system.

Vision AI, a live face, and real-time motor control don't belong on one processor. SafePath splits the work across three boards that talk over dedicated links. Hover the boards — then simulate a fall and watch the alert travel.

ORANGE PI 5 PLUS THE BRAIN — YOLOv8 pose AI cameras · LiDAR · decisions RASPBERRY PI 5 THE FACE — touchscreen UI voice · alerts · "I'm okay" button NUCLEO F446RE THE REFLEXES — motors · mast cliff sensors · real-time control GIGABIT ETHERNET — "fall detected" UART 115200 — drive commands status · mast height
system idle — patrolling

// orange = serial command paths · blue = ethernet — every decision stays on the robot, nothing leaves the home

04 Engineering

Built from the cells up.

Nothing here is a kit. The battery was welded cell by cell, the power system designed rail by rail, and every driver, sensor, and pin mapped by hand.

Power system

PWR-01

A custom 24V lithium pack built from 14 Samsung 30Q cells, spot-welded in a 7S2P layout, protected by a 25A battery-management system with separate charge and discharge paths — every balance junction verified by multimeter before first charge.

pack7S2P · 14× Samsung 30Q 18650
voltage24 V nominal / 29.4 V full
protection25 A BMS + XT60 master disconnect
rails24 V drive · 12 V · 5 V regulated
Photo — battery pack + BMS wiring

Drivetrain

DRV-02

Two 24V brushless motors with hall-sensor feedback drive 8-inch pneumatic wheels through custom 3D-printed hub adapters with heat-set brass inserts — designed in CAD to bridge an 8mm motor shaft to a 6-bolt wheel pattern.

motors2× 57BLDC · 70 W · hall feedback
wheels8" pneumatic · carpet-rated tread
hubscustom printed · M3 brass inserts
controlanalog speed + direction per wheel
Photo — wheel + printed hub adapter

Camera mast

MST-03

A belt-driven 4-stage telescoping slide raises the camera from robot height to roughly 10 feet, powered by a geared motor with an encoder tracking exact height. Try it:

DRAG TO EXTEND THE MAST3.0 ft

Sensing & safety

SNS-04

Three downward infrared cliff sensors stop the robot at stairs and ledges. A 3D LiDAR is planned for mapping and navigation trials, and the battery voltage is monitored so the robot knows when to return and charge.

cliff sensors3× IR photoelectric (stairs/edges)
lidarUnitree L2 3D — navigation phase
visionUSB camera array via powered hub
privacyall AI on-device · no cloud video
Photo — sensor bay / electronics
05 Build log

The honest timeline.

Real engineering is iteration. This is the actual state of the build — what's verified, what's in progress, and what's next.

07 Testing plan

Claims need data.

SafePath's science-fair question: does adding 3D LiDAR to camera-based fall detection measurably improve reliability? The answer comes from controlled trials, not guesses.

PROTOCOL

200+ scripted trials

Repeated fall simulations across varied rooms, lighting levels, camera angles, and occlusion (behind furniture) — plus everyday non-falls like sitting down fast, lying on a couch, and bending to pick things up, to measure false alarms.

COMPARISON

Camera vs. camera + LiDAR

The same trial set runs twice: vision-only, then vision fused with 3D LiDAR. Metrics recorded per trial: detection success, false-positive rate, and fall-to-alert time.

Screenshot — SafePath Alert app
08 SafePath Alert

Family finds out in seconds, not hours.

When a fall is confirmed, the companion app notifies chosen family members with what happened, where in the home, and a verification snapshot — with an escalation path if nobody responds.

push alert + snapshot"I'm okay" cancel on robot escalation timerroom locationon-device privacy