AtlasVLA Remembers the Room With Only a Wrist Camera
Wrist cameras are honest. They show what the gripper sees and forget everything the second the FOV slides past the box.
AtlasVLA (arXiv:2608.06729, submitted Aug 7, 2026) is a vision-language-action stack built to stop that forgetfulness. Instead of mapping the latest frame straight to actions, it keeps a persistent world-ego state: a 4D voxel-hashed spatial memory plus an ego-working memory for task progress. A diffusion transformer (DiT) conditions on both.
Dual memory, one wrist
Persistent World State Memory
- Lift 2D wrist tokens with depth (Depth Anything v3) and hand-eye extrinsics into 3D.
- Fuse into a global voxel-hashed map with confidence-weighted neighborhood updates.
- Keep a sliding temporal window and permanently anchor the first frame (often the best workspace overview).
Ego-Working State Memory
- Learnable intent queries pull task-progress tokens from the VLM stream.
- Redundancy-aware consolidation updates an ego memory bank so multi-step jobs do not lose their place.
Actions come from a step-wise conditioned DiT that attends first to ego context, then to retrieved world state.
Results worth the headline
Strict wrist-only setting (no third-person cam):
| Benchmark | AtlasVLA | Note |
|---|---|---|
| LIBERO average | 97.6% | Outperforms multi-view baselines on the table |
| LIBERO-Long | 94.6% | +7.0 pp vs wrist-only MemoryVLA |
| RLBench (6 tasks) | 70.8% | Wrist-only, 128×128 RGB |
| Real general (6 tasks) | 78.7% avg | 50 trials/task, Franka |
| Real long-horizon (4 tasks) | 69.5% avg | +17.5 pp vs π₀ (52.0%) |
The abstract also cites +9.4% absolute on LIBERO-Long relative to the comparison framing in the paper. Ablations are blunt: drop world state memory and real long-horizon success falls from 69.5% to 54.0%.
A Human’s Take
I’m so here for “the wrist cam is enough if you bother to keep a map.” Multi-camera cells are a luxury; mobile manipulators and humanoids will live with FOVs that move when the arm does. If voxel fusion stays cheap enough at 10 DDIM steps, this is the kind of boring systems work that makes foundation policies usable outside a lab cage with ceiling cameras.