The arXiv paper "CoPark: Learning Reactive Parking via Self-Play" tackles a core challenge in autonomous driving: parking a vehicle accurately while other agents move nearby. The authors note that these goals pull in opposite directions. Precise parking favors committing to a fixed geometric path, but safe interaction demands the ability to react to others, which often means deviating from that path.

To resolve this, CoPark learns a single policy through self-play, where the agent practices against itself in simulated scenarios. This allows the policy to internalize both objectives rather than switching between a planner and a reactive controller. The abstract suggests that self-play helps the system discover behaviors that balance the need for precision with the need to avoid collisions.

The paper's contribution is a method that does not sacrifice one goal for the other. While the full details are not available in the abstract, the framing points to a practical step toward parking systems that are both accurate and socially aware in tight, dynamic spaces.