Search papers, labs, and topics across Lattice.
New York University
4
0
2
Even top-performing autonomous driving policies can falter in critical moments, failing to predict surrounding vehicle movements when it matters most.
EMAgnet reduces exploitability in self-play scenarios by dynamically adapting regularization to the agent's evolving strategy, outperforming traditional methods.
Scaling self-play training directly from pixels leads to competitive autonomous driving performance without human trajectory supervision.
Current traffic simulators are inadequate for testing autonomous vehicles because they lack sophisticated AI models of human drivers and environments.
