# No wrist camera, yet VLA success rate is higher

No wrist camera, yet VLA success rate is higher

Full article: https://haiguangboy.com/posts/foveavla

## Core method

Each query predicts 16 steps of actions and executes the first 8; at the same time it predicts the interaction region for the next query.

↳ Wait for the new image to arrive, crop the local region from the original-resolution image using the previously predicted box, then resize it to 224×224.
↳ The panorama preserves layout, the local crop provides detail; RGB and depth use the same crop.
↳ Actions and boxes share the observation representation, but their output branches are isolated. The box influences subsequent actions by changing the next observation, rather than feeding the current box into the current action head.

Region supervision comes from RoboInter interaction annotations on DROID, with no need to collect human eye-tracking data.

## Key results

Four real-robot task suites, 12 trials each:
· With wrist camera π0.5: 52.1%.
· One external RGB-D view: 70.8%.
· Two external RGB-D views: 77.1%.
Ordinary pick-and-place is 100% in all cases; the gains appear in distractor, cloth, and distant-object tasks.

More important is the ablation: two external RGB views with full images only 25.0%, adding box supervision 31.3%, and actually feeding in local crops 66.7%. Knowing where to look is not enough; the model must actually see more detail.

## Comparison with related lines of work

- Same line · [Agile-WAM: lightweight visuo-tactile joint prediction](https://haiguangboy.com/posts/agile-wam) `agile_wam_an_agile_tactile_world_action_model_for_contact_rich_robot_control_2026_09`: 48 real-robot trials with full success rate cannot be extrapolated to industrial reliability validates small-sample and best-of-selection evaluations cannot be extrapolated to stable general capability
- Same line · [D4RT_Efficiently_Reconstructing_Dynamic_Scenes_One](https://haiguangboy.com/posts/d4rt_efficiently_reconstructing_dynamic_scenes_one) `d4rt_efficiently_reconstructing_dynamic_scenes_one_2026_06`: crops from original resolution while preserving the panorama validates can handle in-the-wild video, continuous query coordinates support arbitrary/sub-pixel resolution decoding
- Same line · [DexTouch-WM: human tactile pretraining](https://haiguangboy.com/posts/dextouch-wm) `dextouch_wm_learning_action_conditioned_tactile_world_models_from_human_touch_fo_2026_09`: learning from human wrist-camera-free video remains future work validates human data is not a direct transfer of hardware-free, robot-data-free data
- Same line · [Introducing Gemini Robotics ER 2](https://haiguangboy.com/posts/google_blog_gemini_robotics_er_2_20260730) `google_blog_gemini_robotics_er_2_20260730_2026_07`: zero-shot leaderboard comparison is not a controlled comparison validates numbers lack controls: competitors not named, baseline values not given, comparison charts not verifiable
- Same line · [MT-WAM: how future supervision serves action](https://haiguangboy.com/posts/mt-wam) `mt_wam_reorienting_the_one_pass_predictive_representation_toward_action_generati_2026_09`: main results are not single-variable experiments that only remove the camera or only add cropping validates causal contributions of supervision target, capacity, and branch structure are not yet fully separated
- Same line · [SmoothRL: only accounting for actions actually executed](https://haiguangboy.com/posts/smoothrl_online_reinforcement_learning_during_asynchronous_execution) `smoothrl_online_reinforcement_learning_during_asynchronous_execution_2026_09`: SmoothRL: only accounting for actions actually executed

## Boundaries

⚠️ One fewer wrist camera does not mean zero cost
On an RTX 5080, a Single action query takes about 125ms and box decoding about 50ms. Box decoding overlaps with action execution, so under the paper's setup there is no need to wait for a new frame before localizing, but the computation does not disappear.

Each method has only 48 real-robot trials, so it cannot be extrapolated to industrial reliability; training on wrist-camera-free human video also remains future work.

## Author's assessment (not part of the paper; cross-paper synthesis)

💡 Putting it back into the technical lineage of prediction and control
MT-WAM reminds us: a prediction task being useful does not mean every predicted token should directly become an action condition. FoveaVLA offers another way to connect them: the prediction result changes the visual detail collected next time.

What is worth testing is whether, under occlusion, camera position changes, and larger workspaces, this "predict—observe—act" loop can keep delivering gains, rather than merely predicting the target box more accurately.