First leg test for a low-cost gimbal motor robot dog

Oct 11, 2026 Last reply: 2 hours ago 4 Replies
Jump to replies (4)

Most legged robot builds reach for hobby servos or expensive brushless actuators, so a dog built from scratch around low-cost gimbal motors is a refreshing starting point. The first milestone is a single leg with 2 DOF, and it already moves with real authority: it bends and extends fast enough to jump its test rig up and down the rails.



The test setup looks tidy. The leg sits on two vertical metal guide rails clamped to a desk, with grey 3D-printed linkages, housings and segments driven by integrated actuators. Wires run to breadboards, power lines and control boards beside a laptop showing the source code. Each actuator combines a planetary gearbox with a belt reduction for roughly 14:1, a mix used to keep the design compact. The motors themselves are 2805 140kv gimbal units.



The control stack is the part that impresses me most. A Raspberry Pi runs the gait-style movement and sends target commands to a master STM32 at 50Hz. That master connects to four slave STM32s, one per leg, each driving three motors with FOC at a 6kHz loop. For such a budget-friendly actuator choice, that is a serious amount of control engineering behind it.



Homing is sensorless at this stage: the legs fold to the home position, and the robot detects it by an increase in error rate. Once powered and homed, it always knows its target position, with slight backlash. Adding encoders to the other joints was rejected since that would raise design complexity, electronics and cost. The next step is a PCB to test all 12 motors, with the finished robot planned at 12 DOF and about 3kg. In the demo the leg also pushes against an open hand placed beneath the foot and retracts, and that push looks convincing to me.


Pairing a planetary gearbox with a belt reduction to reach about 14:1 seems like a clever call to me. Keeping each actuator compact matters when the plan is twelve of them on a robot weighing roughly 3kg. The trade-off, in my view, is that two reduction stages mean more parts to assemble and align correctly in every one of the printed housings.

How does the maker find the home position without encoders on the joints? With sensorless homing I would worry about repeatability between power cycles, so I am curious what signal tells the controller the leg has actually arrived at home.

The maker has described this. The legs fold to the home position, and the home position is detected by an increase in error rate. Once powered and homed, the robot always knows its target position, with slight backlash reported. No encoder feedback is involved in that process at the moment.

While waiting for the PCB that will let all 12 motors run, it could be worth trying the learning-in-simulation route on this single leg first. The maker already plans to compare a classic gait controller with simulation learning for the full 12 joints, and a trained policy could be rerun on the existing rig at no hardware cost, letting both approaches be judged before the full robot exists.

Join the Discussion

Have something to add? Share your thoughts — no account required.

Didn't find your answer?

Ask the community — no account required