X-NavDP lifts hard-case robot navigation to 65%

X-NavDP raised real-world hard-case robot-navigation success from 10% to 65% after reinforcement-learning post-training across different embodiments. The paper, code, and model release reports a parallel gain from 61.20% to 84.28% in simulation, using group-normalized trajectory scores and behavior-preserving exploration to adapt a demonstration-trained diffusion policy. For robotics teams building simulation-first testing funnels, the result puts cross-embodiment RL on the pilot list, but 90 real-world trials across nine robot-and-environment settings still make internal replication essential before deployment.