Embodied AI papers make reliability measurable at execution time

Several papers treat VLA reliability as a runtime problem with concrete failure modes. DEFLECT targets asynchronous inference, where a robot executes an old action chunk while the next one is still being computed. It trains on fresh-versus-stale action preferences and raises Kinetix success to 83.3% across delays d=0–7, with 73.5% success at unseen high delays d=5–7.

RoVLA adds consistency losses for paraphrased instructions, denoising timesteps, and perturbed observations. Its evidence is more qualitative in the available excerpt, but the setup covers LIBERO-Plus perturbations across layout, camera, robot initialization, language, light, background, and sensor noise. RoHIL gives a narrower real-robot case: it relights recorded trajectories and fine-tunes offline, reaching 1.00 source and 1.00 shifted-light USB insertion success in the reported anchored setup.

Training papers in this window focus on which updates matter for closed-loop control. PAPO-VLA separates planning actions from dense execution actions, then increases update weight for actions with estimated causal sufficiency and necessity. The excerpt does not include final PAPO-VLA scores, so the grounded contribution is the action-level advantage design.

Pion addresses a lower-level optimizer failure. It argues that Muon can amplify noisy small singular directions in low-rank VLA action-module gradients and low-signal reinforcement learning with verifiable rewards (RLVR). Pion keeps the Muon-style update path but suppresses tail singular directions. On VLA-Adapter with LIBERO Object, it reaches 100% success after 1,500 steps, compared with 97.0% for Muon and 32.2% for AdamW.

MetaFine is the strongest evaluation paper in the period. It breaks manipulation tasks into language understanding, spatial perception, and motor behavior checks, then replaces a single pass/fail score with atomic skills such as grasp-part, align, insert, press-part, toggle-part, rotate-along, and slide-along.

The reported gaps are large. Conventional evaluation can inflate fine-grained capability by up to 70%. Object-level grasping is often above 95%, while the best policy reaches 80% on Grasp Part, 68% on Press Part, and 12% on Rotate Along under part-level constraints. In semantic substitution tests, all five evaluated VLAs score 0% on the modified instruction. On peg-in-hole, overall success stays near zero, but stage metrics show where policies fail.

World-model work spans robot rollouts, video physics, and synthetic aerial data. WEM splits long-horizon prediction into scene-level world state and robot or object ego state, then tests hybrid navigation-manipulation tasks on HTEWorld. That dataset contains 125K training clips, more than 4.5M frames, 300 evaluation trajectories, and more than 2K instructions.

PhyWorld post-trains Wan2.2-I2V-A14B for video continuation and physics preference. Its results are video metrics, not robot control: 0.769 on VBench and 3.09 on a physical-faithfulness benchmark. FlyMirage applies generative scene creation to aerial vision-language navigation, producing 500 3D Gaussian Splatting scenes and about 50,000 dynamically feasible 6-DoF UAV trajectories at a reported cost of about $2 per scene.