Robot VLA reliability depends on recoverable control and compact foresight

Reliability is defined through what the policy does after the task starts to go wrong. RePO-VLA trains on success, failure, and recovery rollouts with separate labels, then reports average adversarial success rising from 20% to 75% and up to 80% in scaled real-world trials. The key detail is supervision over adverse states, contact drift, and useful failure prefixes, not only clean demonstrations.

ECHO attacks the same long-horizon problem through memory. It stores successful subgoal segments in a hierarchical memory and retrieves them during inference. On LIBERO-Long, it reports 93.5% success versus 80.7% for the vanilla π0 baseline. Together, these papers make recovery and memory measurable parts of VLA execution quality.

World-model work favors small internal state that can still guide actions over long horizons. OneWM-VLA compresses each view into one semantic token per frame and jointly generates future latent tokens and action chunks. It reports 98.1% average success across LIBERO suites and 95.6% on LIBERO-Long, while training only a small LoRA adaptation on a frozen backbone.

ConsisVLA-4D adds compact multi-view 3D perception and future-scene reasoning. Its results claim a 21.6% improvement and 2.3× inference speedup over OpenVLA on LIBERO, plus a 41.5% improvement and 2.4× speedup on real robot platforms. The common design target is a policy that can look ahead without making inference too expensive for control.

Several papers make target binding a first-class control problem. OA-WAM splits each object slot into a fixed identity address and a changing content state, then routes attention through the address slice so the action decoder can keep track of the instructed object. It reports 97.8% average success on LIBERO and stronger geometric generalization on LIBERO-Plus than π0.5 on the reported geometric average.

This matters because scene variation can preserve the visible object while changing layout, camera view, robot start pose, or nearby distractors. ConsisVLA-4D addresses a related issue with instruction-relevant object selection and compact cross-view geometry. The week’s evidence points to object-level state as a practical route for manipulation policies that must survive scene changes.

Open robot policies create a second reliability question: whether a released model can be copied, modified, or audited without changing normal robot behavior. GuardVLA embeds a secret visual watermark during training and checks ownership later with a swapped verification head. In LIBERO experiments, watermarked models keep watermark identification confidence near 100% while clean models stay near zero, and benign task success remains close to clean baselines.

The daily trend also notes ATAAT-style backdoor risk after tuning. That places release safety in the same evaluation frame as control reliability. A deployable VLA now needs task success, recovery behavior, and evidence that the released policy can be checked after downstream adaptation.