Hybrid RL Local Planner
Deployed on a real autonomous vehicle — 80% less cross-track error than a classical planner, under 10 cm at over 10 mph.
- Year
- 2026
- Category
- robotics
- Built at
- Dimaag.ai
Stack
- Problem
- A local planner decides, many times a second, how the vehicle should steer to stay on the path it has been given.
- Approach
- Trained with PPO in NVIDIA Isaac Lab, alongside a 2D LiDAR simulation pipeline built in Isaac Lab and Gazebo.
- Result
- 80% less cross-track error than the classical planner alone.
My summer 2026 work as a robotics intern at Dimaag.ai: a local planner for autonomous-vehicle trajectory tracking that pairs a reinforcement-learning policy with a classical planner, taken from simulation onto a physical vehicle.
The problem
A local planner decides, many times a second, how the vehicle should steer to stay on the path it has been given. The goal was to track complex paths at speed more tightly than the classical planner does on its own — and to do it on the real vehicle, not only in simulation.
Architecture
Trained with PPO in NVIDIA Isaac Lab, alongside a 2D LiDAR simulation pipeline built in Isaac Lab and Gazebo.
Getting it off the simulator. Domain randomization during training, and testing the same policies across simulators, were how the policies were made to transfer to hardware rather than to one simulator's physics.
On the vehicle. The policies were deployed and validated on a physical autonomous vehicle. The work lived in Python modules inside a shared robotics codebase, with Git-based code review and containerized development.
Measured results
80% less cross-track error than the classical planner alone. On the physical vehicle, under 10 cm mean trajectory error at speeds above 10 mph on complex paths.