Yuxin Cai

Predicting compact, learned future representations for world-action modeling slashes training costs by 65x and outperforms native visual targets, even with smaller models.

Embodied agents waste critical step budgets exhaustively clearing dead-end rooms; gating local exploration with explicit LLM-arbitrated relocation more than doubles navigation efficiency on Matterport3D.

Across seven representative adapted baselines, Grounding2Route substantially outperforms existing methods in all metrics, and a substantial gap to human demonstrations remains, highlighting the difficulty of Map2Route and the considerable headroom for future progress.

Roken can generate coordinated multi-robot trajectories in real-time, achieving higher success rates than traditional sequential planning methods.

Navigating with fewer than 8 VLM calls per episode, Goal2Pixel redefines efficiency in vision-language navigation tasks.

By explicitly modeling CAV-CAV vs CAV-HDV interactions, this MARL approach substantially improves traffic flow in mixed autonomy scenarios.