Closed-loop VLA training and evaluation changes

VLA training teams should replace gripper-state-only wrist-camera masking with a gate that also considers the current sub-task and its physical constraints. AC-VLA’s closed-gripper masking discourages visual shortcuts and contributes to large LIBERO-OOD gains, but IMBench shows that alignment, tool use, hidden state, timing, and balancing remain execution bottlenecks even when models identify constraints. Those operations may require close-range or force-sensitive observations after the gripper closes, so one masking rule can trade trajectory memorization for sensor deprivation.

The concrete change is to attach constraint labels to AC-VLA’s automatically segmented sub-tasks: suppress the wrist view during transport when third-person geometry is sufficient, but preserve it for contact, alignment, tool, hidden-state, and stability phases. A useful first check is an inference-time occlusion ablation by phase on held-out LIBERO-OOD and IMBench episodes; if wrist removal disproportionately damages those constraint classes, the training gate should be conditioned on constraint type rather than gripper state alone.

Teams comparing many action experts or tenant-specific VLA policies can reduce duplicated evaluation compute by forking simulator states from a common observation history. Compute the slow backbone and its cache once for the shared prefix, attach each isolated action expert to that cache, and branch the environments only when their actions diverge. JoyNexus already isolates tenant-specific modules and groups compatible workloads around a resident backbone; the fast-slow driving system shows that an action expert can consume a persistent backbone cache plus the current frame at control rate.

The evaluator must preserve closed-loop branching rather than score every expert on one fixed log: the driving study’s route completion rose from 82.1% at 10 Hz to 94.0% at 20 Hz with the same expert, showing that control cadence changes outcomes. The cheapest implementation check is to fork identical simulator snapshots, compare shared-prefix and independently computed caches for numerical and trajectory equivalence, then measure GPU time saved before branches diverge. JoyNexus does not report numerical efficiency gains, so that measurement is necessary before adopting the design.