← All projects

IFAC MECC

Agile Off-Road Traversal of an Ackermann-Steered Platform

Give the vehicle a goal pose in rough terrain and let it work out how to get there — without a hand-designed motion plan in the middle.

86% success rate reaching the desired goal pose

The problem

Point stabilization — get to this position and this heading — is a well-understood problem on flat ground and a nuisance off it. An Ackermann-steered vehicle can’t turn in place, so the approach has to be planned; add slopes and loose surfaces and the plan stops matching what the vehicle can actually execute.

What I built

A learning-based goal-to-goal navigation policy that treats motion planning and control as one problem for a mid-scale off-road platform. The vehicle learns the approach behaviour directly rather than tracking a separately planned trajectory.

Reward shaping was the real work here. It’s easy to write a reward that produces a policy which reaches the goal by driving in a way no one would sign off on — so the shaping deliberately favoured stability over aggressive maneuvers, and I authored a custom off-road simulation environment so robustness was trained rather than assumed.

Learned goal-to-goal navigation on the off-road platform.

Result

An 86% success rate at reaching the desired goal across the evaluation set. Published in IFAC-PapersOnLine through the Modeling, Estimation and Control Conference.

More projects

See all →

Current work

Digital Twins and Edge Autonomy at Arrive AI

An autonomy stack combining RL policies with vision-language-action models, the edge inference that runs it on Jetson hardware, and a digital twin to test it before deployment.