Action learning is getting sharper, while VLA robustness still lags

World-model work in this period is getting more concrete about where predictions fail and how to use that signal for training. WAV treats verification as two simpler checks: whether a future state looks plausible and whether the action could actually reach it. That design is aimed at the sparse-data regime that hurts action-conditioned dynamics models most. The headline numbers are strong for a single-day crop: 2x sample efficiency across nine tasks and an 18% downstream policy gain. The same period also extends world-action modeling into driving. DriveDreamer-Policy predicts depth, future video, and action in one stack, with depth generated first as the geometric scaffold. On Navsim, it reports 89.2 PDMS on v1 and 88.7 EPDMS on v2, alongside better future-video quality.

VLA finetuning papers are focusing on action distributions, not just bigger backbones. The FAN prior paper argues that many robot states admit a small set of near-equivalent good actions, so training should keep probability mass over that neighborhood. In ManiSkill supervised finetuning with OpenVLA, it raises in-distribution success from 78.1 to 89.8 and lifts average OOD success from 58.1 to 63.3. A separate control paper makes a related point lower in the stack: controller gains change how easy policies are to learn and transfer. Its behavior-cloning results favor compliant, overdamped gains, and its retargeting setup keeps at least 90% success with joint-position MSE below 1e-3 across gain settings up to 25x decimation. Taken together, the practical emphasis is clear: action tolerance and control interface choices are becoming first-order training variables.

Embodied VLA work is spreading into new operating domains, with speed and structure built into the model. UAV-Track VLA adds temporal compression, spatial grounding, and a 25-step flow-matching action head for instruction-following drone tracking. It is backed by a large CARLA benchmark with 892,756 frames, 176 tasks, and 85 objects, and reports 61.76% success on long-distance pedestrian tracking with 33.4% lower per-step latency. Driving papers are also organizing VLA systems by role. UniDriveVLA splits understanding, perception, and planning into separate transformer experts to reduce interference between language reasoning and spatial perception. Its excerpt does not expose the final headline row, but it claims state-of-the-art results on nuScenes and Bench2Drive and gives stronger baseline context than earlier driving VLA reports.

Robustness remains a hard limit for current VLA systems. Tex3D shows that adversarial 3D textures on manipulated objects can break policies at rates that are hard to dismiss as edge cases. The attack is physically grounded at the object surface, optimized through a differentiable rendering path, and stabilized across long episodes with trajectory-aware weighting. Reported failure rates rise from 24.1% to 88.1% on OpenVLA, from 4.7% to 76.0% on OpenVLA-OFT, and from 4.6% to 71.8% on pi0 under untargeted attack. On OpenVLA spatial tasks, failure reaches 96.7% under targeted attack. This is a useful counterweight to the period's capability gains: better training and better planning do not remove large visual attack surfaces.