The public research adds official SmolVLA / Panda closed-loop tasks, complete-version comparisons, physical fallback and fair-cost experiments, with recorded successes and failures.
The official policy completes 58/100 tasks in the original configuration. A fresh-state confirmation uses MuJoCo 3.3.7 and predict 50 / execute 10, completing 93/100 while retaining seven timeouts. The runs use separate configurations and state grids; both record zero observer interventions.
The update also compares 600 routing cases, 324 physical stress cases and 60 fresh obstacle-field cases. Completion, rejection, unsafe actions and computational cost are reported separately, retaining the duration cost of slow fallback and the approximately 1 ms mean marginal saving with worse P95 for incremental checks.
Four public recordings show original-configuration success, exact failure replay, low-friction fallback and earlier UR5e mechanism cases. Each is labelled with its study; the project page links to formal records, models and release material.