Robot learning is being judged inside the control loop. The strongest papers add future-change priors, drift monitors, critics, world-model rollouts, and cheaper motion data to make vision-language-action (VLA) policies…
Robot teams can test three concrete changes with the current evidence: interrupt stale action chunks during inference, collect moving-camera episodes alongside static views, and pretrain VLA policies on unlabeled robot…
The period’s strongest signal is practical deployment pressure on Vision-Language-Action (VLA) robot policies. MolmoAct2 makes the reproducibility case with open weights and robot datasets.
Robot manipulation teams can now test VLA claims with more concrete gates: per-step latency under skipped backbone calls, stale-observation success under delayed action chunks, targeted simulation-video augmentation for…