Builders on the hardest parts of building robots

Every robot builder has a story about the thing that took far longer than it should have, so we’re talking with builders to hear them.

We reached out to hobbyists, students, researchers, and engineers who were sharing their builds online. We asked what they’re building, what took far longer than they expected, and what one part of building robots they’d make easier if they could.

While their machines are very different, from a desktop arm to an 18-servo hexapod and a walking robot turned into a pair of robot arms, the problems they ran into centered around the same three things. They struggled to see what was happening inside their hardware. What worked in simulation often fell apart on the real robot, and just finding and setting up the right parts took far longer than it should have.

Here's what they told us, in their own words.

1. You can't see what the hardware is doing

When a robot doesn't do what you expect, it's hard to tell whether the problem is in your code, your wiring, or the parts themselves. Again and again, the problem was hiding somewhere the builder couldn't see.

Chernor Sulaiman Jalloh, a software engineer and AI/software development facilitator, kept watching his robot arm fail, so he assumed his code was wrong. He rewrote parts of it, added telemetry, and even built a browser simulator to test it, which showed everything working perfectly. A friend finally suggested reprogramming the servos, and that fixed it.

“Three days of questioning my code, building an entire simulator, and researching different solutions, only to discover that the issue wasn’t with the software in the first place.”

What he wants is better visibility across his code, firmware, and hardware.

“I’d love to see development tools that provide better visibility into what’s actually happening across those layers.”

Software engineers Antoine Boxho and Finley Warman rebuilt a hexapod from scratch, with a Raspberry Pi 5 running its 18 servos. They found that the obvious cause was rarely the real one.

“Debugging hardware faults. The obvious suspect was usually wrong. An I2C fault I blamed on the power regulator turned out to be a bad crimp and a faulty hub, and a servo issue I put down to protocol was really the mounting angle.”

I2C is the connection that lets the chips in a robot talk to each other, and a crimp is the small metal connector pressed onto the end of a wire.

Some failures, they found, don't show at all until it's too late.

“A servo stalling against a bad mount looks fine until it’s dead.”

Kapalik Khanal, a robotics researcher, explained why hobby servos are so hard to see into. You send them a target position, and a sealed controller inside handles the rest.

“Hobby servos are position-controlled, with a hidden internal controller, backlash and bus latency, and none of that exists in simulation unless you model it yourself.”

2. The sim works, but the real robot doesn't

The problem we heard most was the gap between simulation and the real robot.

Kapalik first built a small bipedal robot eight years ago using basic hobby servos, an Arduino, and inverse kinematics, without any learned control policy. He recently revisited bipedal locomotion with K2, training a walking policy in simulation before transferring it to hardware. When we asked what took far longer than he expected, this was his answer.

“Making K2’s simulation match the real robot. Inertias, joint ordering between sim and hardware, and frame conventions all had to line up exactly, and a single mismatch was enough to make a policy that walked well in sim fail on the real robot.”

A policy is the skill the robot learned in simulation, in this case, how to walk. For it to work on the real robot, how heavy each part is and how that weight is spread, the order the joints are listed in, and which way each axis points all had to match the sim exactly.

Gaurang Solanki, an AI engineer, is teaching a robot arm to pick up a cube in MuJoCo, a popular physics simulator, before moving what it learned to the real arm. Training was the easy part.

“Training runs on its own, but matching the sim to the real arm took hours of manual tweaking.”

Even though he built the sim from the real arm's own design and calibration, the motors behaved differently once they were real.

“They still differ: the real motors are softer and react a few steps late, and the trained policy moves faster than I’m comfortable running the real arm.”

Mamadou Sory Maiga, a mechatronics engineering student specializing in robotics, is still tuning his arm in simulation before he prints parts, and a previous project taught him what to expect.

“Moving from sim to hardware is challenging because capturing accurate inertia values for printed components, backlash, and precise actuator specifications is notoriously difficult.”

Backlash is the small bit of play in a gear that lets a joint wiggle before it moves, and it rarely shows up in a sim.

Kapalik summed up what he'd change if he could.

“If actuators came with accurate simulation models, policies trained in simulation could transfer more reliably to hardware, allowing builders to focus on what the robot does instead of characterizing their motors.”

3. Parts that fight you

Before a robot takes its first step, builders have to find the right parts, set them up, and keep them from breaking, and every one of those took longer than it should have.

For Mamadou, it starts with just finding the right part.

“Finding reliable, affordable parts with exact specifications. It often takes hours to locate a component, only to find that vital parameters are missing or that a reliable option is prohibitively expensive.”

Kapalik took apart his walking robot, K2, and reused its 12 servos to build a pair of robot arms, one he guides by hand and one that copies its moves.

“The mechanical swap was the easy part; the configuration was the real work. Every servo still carried K2’s settings, so I reassigned IDs, reset zero offsets, and recalibrated every joint range.”

Each servo needed a new ID so the controller could tell them apart, and each joint's zero point and range of motion had to be set again for the new robot.

And when a part fails, there's often nothing to do but replace it. Antoine and Finley lost three servos on their hexapod.

“I burnt three of them because the arms were mounted off centre and I had nothing watching the current. Every time it’s a sealed box you just throw away and order again”

What we're taking away

To make progress, these builders often had to build their own simulators, digital twins, cameras or probe tools along the way. That's time spent on the stack instead of on the robot they set out to build.

Hardware that respects your time is what we're building at Springcraft. Our smart components report their health. Every part ships with its own simulation model, and the sim runs the same firmware as the real thing. Plugging in a part means the system knows what it is, what it can do, and how to talk to it, so you can focus on making your robot great.

A huge thank you to the builders who shared their builds and their honest answers with us.

Building a robot? Tell us what’s slowing you down at contact@springcraft.ai.