An electric golf trolley that follows you is not a new idea. The market has a few. They cost $1,500 to $3,000, they use Bluetooth or GPS, and most of them work fine on a flat course with no one around and then lose you the moment a tree gets between you and the trolley.

We have been building industrial auto-following systems for years. The technology to do this properly, with centimeter-level UWB ranging instead of fuzzy Bluetooth distance, exists and is affordable. So we wrote this guide for anyone who wants to build their own. Whether you use our modules or source everything yourself, the architecture is the same.

What a following golf cart actually needs

Four things, working together:

1. A tag you carry. Small, battery-powered, goes in your pocket or clips to your belt. It sends UWB pulses the cart can measure.

2. Anchors on the cart. Two UWB arrays that receive those pulses and measure both distance (TWR) and angle (PDOA). One faces forward, one faces aft, so the cart knows where you are through a full 360 degrees.

3. A controller. The brain. It reads the UWB data, runs the pursuit loop, reads the lidar, runs the avoidance layer, and sends commands to the motor driver.

4. Motors and a lidar. Two drive motors for the wheels, plus a 360-degree single-line lidar for obstacle detection. Trees, other players, golf bags left on the path.

Choosing a frame

You can start with an existing push trolley or build one from scratch. The frame needs to hold the motors rigidly, because flex in the mount translates to bearing noise at the UWB anchors. A wobbling frame means a wobbling distance reading.

If you are modifying an existing trolley, the two things to check are wheelbase width and ground clearance. A narrow wheelbase makes the trolley skittish at speed. You want at least 35 cm between wheel centers for stable steering. Ground clearance needs to clear the lidar housing and the motor pod by at least 8 cm, or the lidar will read the ground as an obstacle.

Mounting the UWB system

The front anchor array goes on the handlebar end of the trolley, facing forward. The rear anchor goes on the opposite end, facing back. Both need clear line of sight in their respective directions. Metal frame tubing between the anchor and the tag will degrade the signal.

The dual-anchor layout solves the 180-degree front-back ambiguity that single-array PDOA suffers from. When you are in front, the front array measures your bearing. When you walk behind the trolley (or it turns past you), the rear array takes over. The VCU fuses both readings with the distance measurement to triangulate your position through a full 360 degrees.

Anchor spacing matters for PDOA accuracy. The two antennas in each array need to be at a specific distance apart, determined by the UWB wavelength. Too close and the phase difference is too small to measure. Too far and you get phase wrapping. Follow the module manufacturer’s spec.

Wiring the motors

A golf trolley with a loaded bag weighs 25 to 40 kg. On a slope, the motors need to push that uphill at walking speed. That means you need motors rated for at least 80 W per wheel, with encoders for closed-loop speed control.

The motor driver connects to the VCU. The VCU sends speed and turn commands, the driver translates those into PWM signals for each motor. For a differential drive, left and right wheels get different speeds: v_left = v - turn_rate * wheelbase / 2, v_right = v + turn_rate * wheelbase / 2.

Use a battery with enough headroom. A 36V 10Ah pack will run a 160W system for 3 to 4 hours of continuous walking, which covers two rounds of golf easily. Budget for a BMS that handles regenerative braking, because the pursuit loop means the trolley stops and starts constantly.

The software: pursuit and avoidance

Two loops run in parallel on the VCU, both at 20 Hz.

The pursuit loop reads UWB distance and bearing, runs a spring-damper controller, and outputs a target speed and turn rate. The details of that math are in our follow-me math article. The short version: a spring pulls the trolley toward the target distance, a damper smooths the motion, and a low-pass filter adds about 250 ms of lag so it feels natural rather than twitchy.

The avoidance layer reads the lidar scan and decides whether the pursuit loop’s chosen path is safe. In open terrain it is mostly passive. When it detects an obstacle in the trolley’s path, it overrides the heading and steers around or brakes. We cover the details in our lidar avoidance article.

Testing: start on flat ground

Do not take it to the golf course on day one. Start in a parking lot or a park with flat, open ground and no obstacles. Tape out a 20-meter straight line and walk it.

The first thing to tune is the follow distance. Start conservative: 2 meters. Walk forward at a normal pace, stop, walk again. The trolley should start a quarter second after you, accelerate smoothly, and coast to a stop about 2 meters behind. If it lunges, lower Kp. If it barely moves, raise it. If it oscillates, raise Kd.

Once straight-line following feels right, test turns. Walk a gentle S-curve. The trolley should track your heading through the turn. If it overshoots and swings wide, increase the turn gain or reduce max speed during turns.

Only after flat-ground following is solid, bring in the lidar. Walk toward a trash can or a bench. The trolley should slow and steer around it. If it stops dead, the obstacle threshold is too aggressive. If it runs into things, the lidar is mounted wrong or the avoidance layer is not triggering.

The golf course is the final exam, not the first lesson. Get it working in a parking lot first.

What goes wrong on a golf course

A golf course introduces three things your parking lot tests will not: slopes, trees, and other people.

Slopes change the motor load. The pursuit loop compensates automatically because it is distance-based, not speed-based. But on a steep downhill the trolley will accelerate toward you faster than on flat ground. Make sure max_decel is set high enough to stop in time.

Trees are a UWB multipath problem. A pulse bouncing off a trunk arrives slightly later than the direct signal. If the reflection is strong enough, the ranging reads the wrong distance and the trolley veers. Our outlier rejection (frames where distance jumps more than 0.2 m get discarded) handles most of this, but a dense tree line can still cause a few bad readings in a row.

Other people are a lidar problem. The avoidance layer sees them as obstacles and steers around them. This is correct behavior. But if someone walks between you and the trolley, the UWB signal gets body-blocked. On channel 9, a human body directly between the tag and the anchors nearly cuts the line-of-sight path. The trolley will lose you for a moment and then pick you back up when the path clears. It is the hardest problem in UWB following and we have not fully solved it yet.

Rough cost breakdown

If you are sourcing everything yourself, expect to spend in this range:

Frame: existing push trolley ($150 to $400) or custom build ($200 to $500 in materials)
Motors + encoders: $80 to $200 per pair
Motor driver: $40 to $120
Battery + BMS: $80 to $200
UWB modules (tag + dual anchors): $100 to $300
Lidar (360 single-line): $100 to $500
Controller: $30 to $150

That is $580 to $1,870 depending on component quality. A commercial following trolley costs $1,500 to $3,000, so building your own can be cheaper, and you get to choose exactly what goes in it.

If you want to skip the component sourcing, the followwagon kit bundles the UWB module set, VCU, and motor driver as one integrated package. You add the frame and the motors. The three-stage roadmap lets you start with our chassis drawings and recommended motors, then move to your own frame and motor choices as you get comfortable.

What to expect when you first build one

The first time it follows you is a strange feeling. You walk and something trails behind you, matching your pace, slowing when you slow, stopping when you stop. It takes about five minutes to stop looking back at it.

But getting there takes patience. The pursuit loop tuning alone took us three months of walking the same path, logging data, adjusting two numbers, walking again. The avoidance layer took another two months of driving into things and figuring out why.

The point of this guide is to shortcut that. The architecture works. The math works. If you build it right, it will follow you. The hard part is the tuning, and that is where the detailed articles and the build notes in each one come in.

If you build one, tell us. We want to see it.