AtlasVLA: Persistent World-Ego State Modeling for Vision-Language-Action Models

The Problem With Current Robot Vision Models

Vision-Language-Action (VLA) models drive modern robot arms, but they're fundamentally short-sighted. They process only what a wrist-mounted camera sees right now—the moment an object leaves the frame, it vanishes from the robot's memory. Over long, multi-step tasks, this creates two critical failures: the robot forgets what objects exist where (perception forgetting), and it loses track of progress toward the goal (temporal forgetting). Current models are purely reactive, responding to immediate observations rather than reasoning about the broader scene. AtlasVLA solves this by giving robots persistent spatial memory.

How It Works

The key innovation is a dual-memory architecture that builds and maintains a coherent model of the world around the robot.

Persistent World State Memory converts fleeting 2D camera images into a lasting 4D spatial map. The system uses voxel hashing—dividing the workspace into small 3D cells—to store observations as they accumulate over time. When the wrist camera pans away from an object, that object's location remains encoded in the world map. Think of it like a 3D occupancy grid that the robot continuously updates as it moves and observes.

Ego-Working State Memory tracks the robot's own movement history and task progress. It records where the robot has been and what steps it's completed, preventing amnesia during long-horizon tasks.

These two memory systems feed into a diffusion transformer (DiT)—a neural network that generates robot actions by iteratively refining noisy predictions into coherent motion commands. By conditioning the action generator on both the persistent world state and ego state, the model can reason spatially and temporally. It knows not just what the camera currently shows, but what exists beyond the frame and how far through the task sequence it has progressed.

Why It Matters

The results are striking. On LIBERO-Long (a long-horizon manipulation benchmark), AtlasVLA achieves a 9.4% absolute improvement in success rate compared to prior VLA methods. In real-world tasks, the gap widens to 17.5%—a decisive margin for practical robotics. Most impressively, AtlasVLA does this using only a wrist camera, decisively outperforming baselines that get data from multiple fixed cameras. This matters because multi-camera setups are expensive, inflexible, and require careful calibration; a wrist camera is cheap and travels with the robot.

The persistent world memory addresses a fundamental gap in embodied AI: robots don't need to see everything at once if they can remember what they've already seen. By decoupling perception from action, AtlasVLA enables the kind of spatial reasoning and long-term planning that humans perform effortlessly. The framework shifts VLA models from reactive puppets into agents that model their environment persistently, opening a path toward more reliable, general-purpose robotic manipulation.