Research questionHow can vision-language-action systems reliably execute long-horizon manipulation while tracking state and conditional dependencies?Short manipulation skills do not ensure reliable completion of multi-step procedures. The robot must retain task progress, respect action dependencies, respond to visual conditions, and ground each action to the correct objects and destinations. Latest papersRecent research connected to this question, newest first.What Matters, When? Diagnosing and Improving Conditional Visual Grounding in Visuomotor Imitation PoliciesEvidence covers Action Chunking with Transformers policies tested in simulation and on a physical UR3e, plus a pretrained vision-language-action policy in state-conditioned instrument handling. Distractors vary in controlled color and shape similarity, with failures examined during picking and placement.research paper · Sep 4, 2026Towards Neuro-Symbolic Procedural Reasoning for Long-Horizon Vision-Language-Action ManipulationThe evidence covers workspace clearing and surgical-instrument handling with learned VLA control, explicit task graphs, multimodal procedural memory, and pseudo-gaze annotations from robot-view teleoperation videos. The initial study bypasses cross-view gaze transfer and evaluates selection, subtask completion, step order, task success, and procedural or execution mistakes.research paper · Sep 4, 2026