Robot VLA reliability is being measured in rollouts, calibration, and safety cost

Several papers treat the deployed robot as a system that needs local evidence before or during the task. ICWM lets a policy run a short probing phase, then uses the observed action-to-image changes as context for control under new camera views or body setups. FORCE uses online rollouts to fine-tune a VLA policy with a calibrated critic and reports real-world Franka success rising from 45.0% under behavior cloning to 98.3% after fine-tuning, with no human intervention during the online stage. PhysReflect-VLA adds execution-time feasibility checks and correction after observed state mismatches, giving smaller but consistent real-robot gains on long-horizon tasks.

Geometry is being added as an execution signal, not only as a visual feature. Reflective VLA stores observation-action-consequence triplets so the policy can infer camera calibration, actuation bias, and robot setup effects at inference time. G3VLA injects camera intrinsics and extrinsics into visual tokens through ray embeddings, projective positional encoding, and cross-view fusion while keeping the base action pathway. Both lines point to the same practical need: a robot policy has to know how its own actions change the scene under the current camera and embodiment.

Safety work is getting more diagnostic. LIBERO-Safety tests whether VLA policies can complete manipulation without unsafe contact or unsafe instruction following, across 75 tasks and 19,664 screened collision-free demonstrations. ForesightSafety-VLA scores each rollout as safe success, unsafe success, safe failure, or unsafe failure, then adds process metrics such as cumulative safety cost and risk exposure time. The important evidence is that task success can hide risky motion: ForesightSafety-VLA reports unsafe success even for the strongest listed baseline, while LIBERO-Safety shows large drops on harder physical-safety levels.

The week also shows more work on making action supervision cheaper and more reusable. InSight segments existing demonstrations into named primitives, lets a vision-language model choose missing primitive attempts, and folds successful robot rollouts back into the VLA. It reports 92% twist success and 96% pour success on real xArm tasks after adding 20 successful acquired primitive episodes for each skill. LA4VLA takes a data-first route: it derives 33,116 language-action episodes from DROID and pretrains policies on instruction-conditioned motion without images, improving a 1B-parameter VLA on simulated and real-world tasks.