AtlasVLA gives robots persistent memory, beating multi-view baselines by 17.5%

AtlasVLA, developed by Guiyu Zhao and 9 co-authors, tackles a core weakness in vision-language-action (VLA) models: their reactive, memory-less behavior. When a robot relies solely on a wrist-mounted camera, objects leaving the field of view cause perception forgetting, and long multi-step tasks trigger temporal task-progress drift. AtlasVLA solves this by building a persistent world-ego state through a dual-memory architecture. A 4D persistent world state memory lifts transient 2D observations into a globally updated, voxel-hashed spatial representation, eliminating visual blind spots. A separate ego-working state memory tracks historical robot states and task progression. Both memories condition a diffusion transformer (DiT), allowing proactive reasoning instead of purely reactive control.

The results are striking: AtlasVLA achieves state-of-the-art performance across LIBERO, RLBench, and real-world benchmarks while using only a single wrist camera. It decisively outperforms multi-view baselines, posting absolute success rate gains of 9.4% on LIBERO-Long and 17.5% on real-world long-horizon tasks. This suggests that dense memory of the environment and task state can compensate for limited sensory input, potentially reducing hardware costs in robotic manipulation. For embodied AI developers, AtlasVLA points toward systems that reason about what they've seen and done, rather than only reacting to the current frame—a significant step toward practical, long-horizon robot autonomy.