Carlos S. Sepúlveda

GRPO eliminates the structural vulnerabilities of baseline-dependent training, achieving robust performance in neural combinatorial optimization tasks without the need for a frozen policy.

Underreported implementation details in classical coverage path planning, like endpoint-aware residual degree calculation, can dramatically impact performance on sparse geometric graphs with bottlenecks.

Ditch the critic: a novel critic-free DRL approach more than doubles the success rate of maritime coverage path planning compared to heuristics, while also generating shorter, smoother paths.