See like a Robot: Robot-Centric Pointmaps for Vision-Language-Action Models

TL;DR — Give a VLA per-pixel 3D in the robot frame,
and it stays robust as training-time camera viewpoint variation grows.

Robot datasets mix demonstrations from many camera viewpoints; as that training-time variation grows,
an RGB-only policy degrades while pointmaps keep it robust, widening the gap to +10.9 points.
The whole method is one extra encoder and one element-wise addition.

Motivation

Large-scale robot datasets aggregate demonstrations collected across labs and episodes, often from different camera viewpoints. DROID, for example, contains 1,417 unique third-person viewpoints.

Why does this viewpoint variation matter? A VLA observes the scene in the camera frame but predicts actions in a robot-centric frame, creating an observation-to-action frame mismatch. With a fixed camera, this mapping remains consistent across demonstrations. As training data spans more viewpoints, however, the policy must generalize the mapping from multiple camera frames to the shared robot-frame action space, making the task harder to learn.

To test this, we conduct a controlled study on 24 RoboCasa tasks. Using a π-style model initialized from a base PaliGemma checkpoint, we vary only the amount of camera viewpoint randomization across training demonstrations while keeping the remaining training and evaluation settings fixed.

−9.6 points

RGB-only degrades as training-time viewpoint variation increases.

Solution

Rather than asking the policy to learn how each camera frame maps to robot-frame actions, we express the observed 3D geometry from different viewpoints in a shared robot frame. This gives the policy a consistent geometric coordinate system across diverse camera viewpoints. Robot-centric pointmaps bridge this frame mismatch while preserving the image-form representation used by pretrained VLAs.

• Robot-frame geometry. Each pixel stores the 3D coordinate of its corresponding scene point in the frame where robot actions are defined.
• Image-form structure. The dense H × W grid is preserved, maintaining spatial correspondence with RGB.

A robot-centric pointmap stores robot-frame XYZ coordinates at the corresponding image pixels, preserving the same dense H × W grid as RGB.

The whole method: one extra encoder and one element-wise addition.

Results

Adding pointmaps to the same controlled study changes the trend: RGB + Pointmap drops by only 1.8 points, compared with 9.6 points for RGB-only.

−9.6 points

RGB-only degrades as training-time viewpoint variation increases.

+3.1 → +10.9 points

The pointmap advantage over RGB-only widens as training-time viewpoint variation increases.

Pointmaps Improve Both Pretrained VLA Backbones

On RoboCasa, third-person camera viewpoints are randomized across training demonstrations. In this setting, pointmaps improve both pretrained backbones: +7.6 on π_0.5 and +4.2 on SmolVLA.

With π_0.5, pointmaps reach 62.9% average success, outperforming all camera-aware, 3D-augmented, and point-cloud baselines.

FP3 is a DROID-pretrained point-cloud policy without a VLA backbone. Per-task-category results are in the paper.

On a Real Robot, the Gap Widens at an Unseen Camera Viewpoint

We collect 180 demonstrations over four tasks on a FR3 while repositioning the external camera across three training viewpoints. We then evaluate the policies at one seen viewpoint and at a held-out viewpoint not observed during training.

At evaluation, the external camera is placed either at one of the training viewpoints (seen) or at a held-out viewpoint not used for data collection (unseen). Pointmaps improve performance at both camera viewpoints, but the advantage grows from +5.0 points at the seen viewpoint to +11.7 points at the unseen viewpoint. By expressing observed geometry in the robot frame, pointmaps provide a more consistent spatial representation when the external camera moves beyond the training viewpoints.

Per-Task Videos, From Training Views to Unseen-Viewpoint Rollouts

The top row shows one training demonstration from each training viewpoint, seen from the policy's external input camera. The rows below show evaluation rollouts at the unseen viewpoint, recorded from a separate fixed camera, comparing π_0.5 and π_0.5 + Pointmap. All clips play at 2× speed.

Ablation Studies

Why does the design look the way it does? Each choice is backed by its own controlled ablation on RoboCasa. The backbone and training recipe stay fixed, and only the 3D input changes.

For these controlled studies, we use a model without robot pretraining to isolate the effect of each input design. Its success rates are therefore lower than those of the pretrained VLAs evaluated later.

1Pre-compute Robot-Frame Geometry▼

Plücker rays describe the camera geometry, and depth provides per-pixel distance. RGB + Plücker + Depth therefore contains the information needed to obtain robot-frame geometry, but leaves it for the policy to infer. Providing robot-frame geometry directly as a pointmap improves success from 31.6% to 34.7%.

Input	Robot-frame geometry	SR
RGB	not provided	27.9
RGB + Plücker	not provided	28.7
RGB + Plücker + Depth	left to the policy	31.6
RGB + Pointmap	provided directly	34.7

2Preserve the Image Grid▼

Pointmaps outperform point-cloud inputs while preserving the dense image grid used by the VLA. Because RGB and pointmap tokens share the same spatial layout, element-wise addition further improves success from 30.7% to 34.7%.

3D input	Encoder	Fusion	SR
None	–	–	27.9
Point cloud	MLP	concat	24.2
Point cloud	PTv3	concat	32.8
Pointmap	image encoder	concat	30.7
Pointmap	image encoder	add	34.7

3Center Geometry at the End Effector▼

Centering the pointmap at the current end-effector position reduces variation caused by absolute workspace location, making similar interactions appear in a more consistent local coordinate system. It improves success under both fixed and randomized evaluation viewpoints.

Pointmap origin	Fixed eval.	Randomized eval.	Drop
Robot base	34.7	32.7	−2.0
End effector	36.9	36.6	−0.3

BibTeX

@article{lee2026seelikearobot,
  title         = {See like a Robot: Robot-Centric Pointmaps for Vision-Language-Action Models},
  author        = {Lee, Byungkun and Hwang, Dongyoon and Kim, Dongjin and Lee, Hojoon and Park, Minho and Choo, Jaegul},
  journal       = {arXiv preprint arXiv:2607.11498},
  year          = {2026},
  eprint        = {2607.11498},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO}
}