Robot action systems are being built around reusable control infrastructure

Video generation is showing up as a planning module for manipulation, not just a data source. Veo-Act uses Veo-3 to predict a future motion sequence, then hands control to a low-level vision-language-action policy during contact-heavy interaction. The reported gains are large on ambiguous scenes and dexterous execution: average success rises from 45% to 80% across the tested sim and real settings, and real-world pass-by interaction improves from 2/13 to 11/13. The paper also makes the limit clear. Video prediction alone can sketch the task, but precise control still needs a reactive action policy.

Several papers focus on the action stack around VLA models. StarVLA packages multiple action heads, backbones, training recipes, and benchmark interfaces under one codebase, with support for seven integrated benchmarks and both vision-language and world-model backbones. E-VLA attacks a different bottleneck: bad sensing. It adds event-camera input so a VLA policy can keep acting in low light and blur. On Pick-Place, the image-only baseline falls to 0% at 25 and 20 lux, while the event adapter reaches 90% at both levels. Together, these papers put attention on reusable policy plumbing and on perception that survives real capture failures.

Infrastructure is also moving closer to deployment. RoboPlayground turns natural-language task requests into executable evaluation tasks with validation and repair, and reports better usability than Cursor and GenSim in a 26-person study. ROSClaw targets heterogeneous multi-robot execution with tool calling, simulation-based safety checks, and shared execution memory. Its evidence is more system-demo than benchmark-driven, but the point is clear: authoring tasks, checking feasibility, and coordinating hardware are being treated as first-class research problems rather than setup work.

Classic reinforcement learning is still active in robot control when the claim is concrete about speed and scale. FlashSAC argues that off-policy RL can stay stable in high-dimensional control if throughput is high and critic updates are tightly constrained. The paper covers more than 60 tasks across 10 simulators and reports the strongest gains on dexterous manipulation and humanoid locomotion, plus a sim-to-real humanoid result where training time drops from hours to minutes. A separate neuro-symbolic VLA report pushes a different efficiency claim, with 95% Tower of Hanoi success and a 100× training-energy reduction, but its evidence comes from a narrow planning benchmark rather than a broad suite.