Robot VLA research is tightening the control loop around action quality and timing

Several papers attack the action stream itself. RotVLA encodes frame transitions as continuous SO(n) latent actions and composes them during training. It reports 98.2% average success on LIBERO and 89.6% / 88.5% on RoboTwin2.0 clean and randomized settings after pretraining on more than 1700 hours of robot and human video.

FrameSkip and AttenA+ make a related claim at the data and loss levels. FrameSkip keeps 20% of unique trajectory frames, with priority for alignment, contact, grasp closure, and release. Its macro-average success across RoboCasa-GR1, SimplerEnv, and LIBERO rises from 66.50% to 76.15%. AttenA+ gives more loss weight to slow, precision-heavy actions and lifts OpenVLA-OFT on LIBERO from 97.10% to 98.60%.

Long-horizon manipulation papers give the high-level planner a narrower contract with the robot policy. VLAs-as-Tools uses a vision-language model planner to call bounded VLA tool families such as grasp, open, or place. Tool-Aligned Post-Training trains those policies on the same invocation unit used at test time, including tool family, local instruction, actions, and progress feedback.

The measured gains are large on harder task suites. The tool-family setup raises π0.5 from 92.4% to 97.2% on LIBERO-Long and from 39.4% to 62.5% on RoboTwin. GTA-VLA adds a human-facing route for correction: the user can provide a point, box, or trace, and the model folds that spatial cue into reasoning before action. It reports 81.2% success on in-domain SimplerEnv WidowX, while its OOD improvement claims lack exact numbers in the available excerpt.

PAIR-VLA makes visual reliability a behavior-level reinforcement learning target. During PPO fine-tuning, it compares action distributions across paired visual variants. Distractor and texture changes should preserve the action distribution; target-pose changes should separate it.

The result is a clear OOD manipulation gain on ManiSkill3 pick-and-place. OpenVLA improves from 77.90% with PPO to 87.00% with PAIR-VLA across table texture, lighting, target pose, and clutter tests. π0.5 improves from 46.25% to 62.87%. The auxiliary losses are used only during training, so the deployment policy keeps the same inference architecture.

Latency work is now tied to robot-visible effects. Realtime-VLA FLASH uses a small draft model and main-model verification to avoid many full diffusion replanning calls. On LIBERO, FLASH with Triton reduces average task-level latency from 58.0 ms to 19.1 ms with a 0.3 percentage-point success drop, and the paper reports conveyor-belt grasping at speeds where the compared methods fail.

D-VLA handles the training side of the same pressure point. It separates rollout traffic from weight updates and pipelines sampling, receiving, training, and distribution. On OpenVLA-OFT, it reports 156.0 steps/s in a single-node setting compared with 108.24 for RLinf-co and 110.88 for RL-VLA³. The excerpt gives throughput numbers, while final success-rate values are not provided.