Yingxue Zhang

Flow-matching manipulation policies can outperform dedicated 3D-conditioned architectures simply by cross-attending robot-grounded spatial tokens into the action expert, beating the 2025 BEHAVIOR-1K challenge winner without modifying host perceptual pathways.

This work proposes StageGuard, an agentic distillation framework for accurate and efficient stage-transition decisions, which combines teacher-model reasoning with demonstration trajectories to generate structured explanations of subtask completion and policy switching.