top of page

Model-Free Reinforcement Learning for Projectile Targeting with Inherited Platform Velocity from a Moving Omnidirectional Robot

Aditya Rastogi
21/11/26

Projectile targeting from a moving robot is more difficult than targeting from a stationary platform because a launched projectile inherits the translational velocity of the robot. This study investigated whether a model-free reinforcement-learning controller could learn to compensate for this effect while controlling a simulated turret-mounted launcher. A Proximal Policy Optimization (PPO) model was developed in Unity using ML-Agents. The controller used eight observations comprising target geometry, robot chassis velocity, turret state, and aiming progress to generate continuous yaw, pitch, and flywheel-speed commands. The training was done with multiple independent seeds at chassis speed in range 0-1.5 m/s and bearing randomized across 3600 for 6 million training steps achieving high training performance (>99% final hit rate). Frozen-policy evaluation was subsequently conducted over 1,000 episodes at each of stipulated chassis speeds. The controller achieved interception accuracies of >99% for stationary chassis as well as for speed up to 0.5 m/s moving towards and 1.0m/s moving away from target. The accuracy degraded to 85.9% for speed of 1.5 m/s moving towards target, reflecting difficulty in acquiring target from closer distances when platform is also moving at higher speed. The evaluation data was benchmarked against inverse ballistics baseline. These results demonstrate high interception accuracy across the evaluated moving-platform conditions.

bottom of page