Tuning Part 2 – Understanding the Steering System
To get the tractor driving as accurately as possible, accurate position sensing is not enough. Even if your GPS sensors were always 100% correct, the tractor could still overshoot and even wobble back and forth if the software does not account for how the rest of the system behaves.
In my case, the steering system needed significant retuning and fixing. These were the main challenges:
Gearbox slop. The steering gearbox had significant backlash, so the front wheels would not turn until the motor had moved through a small deadband.
Delay. The motor controllers responded to a command roughly 200 ms after it was sent.
A non-linear response. To prevent overcurrents, we switched from speed commands to current (torque) commands. The trade-off is a non-linear response with a large deadzone between 0% and 20% command.
These issues sound problematic, but as long as they are modelled, software can be written to manage them. Issues like these are expected: we are keeping the cost of the tractor reasonable (unlike some other autonomy solutions), and I want the code to work on as many different mechanical systems as possible, so I have to assume these issues will occur.
Annoyingly, this meant applying a topic I learned in university: Signals and Systems.
The steering is now described by a model with a dead time, a gain, a deadband and its dynamic response. The model lives in a configuration file, not in the code, and three parts of the software use it. The simulator uses it to make the simulated tractor steer like the real one. The steering position controller uses it to compensate for the deadband. The path follower uses the delay to decide how far ahead to look, which is the subject of the next post. Moving to a tractor with a different steering system means measuring that system and updating the file.
![]() | ![]() |




Comments