Robot VLA claims now need real execution evidence

VLA work is being judged by what happens during execution: blur, lighting changes, delays, perturbations, and fine task stages. StableVLA targets this directly with an Information Bottleneck Adapter that filters noisy visual channels without extra robot data. The reported gains are concrete: 82.0% versus 58.5% on LIBERO Spatial under severity-5 corruptions, and 50% versus 20% on a real Pack Doll task against the VLA-Adapter baseline. The daily trend evidence also points to latency and illumination tests as common reliability checks across the week.

Several papers add explicit geometry to the policy path, not just to perception. GaussianDream trains a VLA policy with 3D Gaussian reconstruction and short-horizon future prediction, then keeps only learned prefix tokens at inference. It reports 98.4% average success on LIBERO, 52.6% on RoboCasa Human-50, and 50.0% in real-robot evaluation. PointACT feeds hierarchical point-cloud features into the action decoder, reaching 96.0% average success on LIBERO and beating SpatialVLA by 17.9 points in the reported table. These results make 3D structure part of the action interface.

The strongest grounding papers add intermediate state that the robot can act on. AVP makes the vision-language model output visual primitive tokens before action prediction, then uses a flow-matching action expert. On Chinese chess manipulation, it reports 90.28% average success versus 62.67% for π₀.₅, while running at 0.27 seconds per instruction. SOMA adds persistent spatial memory for objects outside the current camera view. On five real-world out-of-vision tasks, full SOMA reaches 28.3% average success and cuts time-to-grasp across tasks compared with GR00T-N1.5. The shared point is practical: policies need target state, scene memory, and fast action checks during execution.

Evaluation work is trying to make physical robot testing easier to repeat. VLA-REPLICA uses off-the-shelf hardware, fixed lighting, camera alignment tools, reference placements, and 500 expert demonstrations across 10 manipulation tasks. The full setup is reported at about $1050 and was assembled by a new user in under an hour. Its 90 test scenes cover in-distribution and out-of-distribution evaluation, and π₀.₅ reaches the best reported in-distribution average success rate at 0.54. This type of benchmark lowers the cost of checking whether a policy works outside a simulator.