Robot VLA work is being judged by control latency, action structure, and data reuse

Two manipulation papers treat action structure as a control problem. Libra-VLA splits prediction into discrete macro-direction intent and continuous fine control, then runs its heavier planner less often through an intent buffer. It reports 97.2% average success on LIBERO and 79.5% zero-shot success on LIBERO-Plus.

CF-VLA applies a similar coarse-then-fine idea to flow-based action sampling. It uses two function evaluations: one coarse action-aware start point and one local correction. The reported trade-off is concrete: 96.5 average LIBERO success at NFE=2, 75.4% lower action sampling latency, and 83.0 average success in real-robot experiments.

The strongest data result comes from MoT-HRA. It builds HA-2.2M, a 2.2M-episode human manipulation dataset, then separates intent into a 3D trajectory expert, a MANO-style hand-motion expert, and a robot action expert. On SimplerEnv-WidowX, it reports 66.1% average success, ahead of ThinkACT, SpatialVLA, OpenVLA-OFT, RoboVLMs, and pi0 variants listed in the paper.

M²-VLA addresses generalization inside the model. It keeps the vision-language backbone frozen, selects useful layer features, and retrieves stored meta-skills to guide action prediction. The reported gains are clearest under instruction and object variation: 66.2% success on rephrased LIBERO Spatial instructions and 34.4% on a novel-object pick-and-place test, with smaller drops than OpenVLA and VLA-Adapter.

Deployment work is measuring the whole observe-infer-act loop. The cross-XPU VLA study profiles inference on RTX 4090, Jetson Thor, AGX Orin, Intel B60 Pro, and Ascend NPUs, then ranks hardware by cost, energy, and time. Its pi0 measurements show why peak speed is insufficient: RTX 4090 reaches 102.3 ms, Jetson Thor 246.0 ms, and AGX Orin 920.6 ms before compilation, with different energy costs.

AsyncShield handles another timing failure mode: stale cloud VLA waypoints. It realigns delayed waypoints with an SE(2) pose transform and uses a safety-constrained local policy with LiDAR input. Under mixed network degradation, it reports 76.7% success, while the no-temporal-alignment ablation falls to 36.7%.