← News

Models & toolsSources: arXiv · 19 Nov 2026

WACO re-anchors few-shot robot policies when cameras or robot bases move

Researchers present WACO, a framework that lets vision-language-action models tolerate shifted cameras and robot bases using only a single RGB-D frame, plus a new pose-shift benchmark.

In plain words

Robots trained on a handful of demonstrations often fail when their camera is bumped or the robot base is moved slightly, because the policy memorised the original setup. WACO uses the robot arm itself as a measuring reference to estimate how things shifted, then rewrites the camera image and robot commands into the original coordinates before further training. The authors report large success-rate gains without needing extra calibration data.

What was announced

The arXiv paper argues that both failure modes are one phenomenon: the policy has memorised a fixed camera-to-base extrinsic, meaning the spatial relationship between camera and robot. WACO places the arm, via forward kinematics at measured joint angles, as a metric calibration target, while the static scene acts as a world anchor distinguishing a moved camera from a moved robot. From one RGB-D frame it estimates both shifts without ground-truth extrinsics or extra data, canonicalises the image and the state, action and start pose to the calibration frames, and fine-tunes the unmodified policy under the artefacts of its own estimator.

The team also introduces LIBERO-PoseShift, described as the first benchmark for RGB VLAs with graded camera, base and compound shifts. Across that benchmark, the LIBERO-Plus camera track and four real-world tasks on a UR5e robot, the authors report that WACO consistently outperforms the standard few-shot recipe and prior viewpoint-robust methods, raising shifted success by 13.2 points in simulation and 36.7 points on the real robot, recovering 90% of the simulation gap and 87% of the real-robot gap to a ground-truth oracle, with no test-time calibration.

What it does not show

  • Evidence is limited to the paper's abstract; experimental details, baselines and real-world conditions cannot be verified from the excerpt.
  • The reported gains are the authors' own claims and have not been independently confirmed.

For security and IT teams

Scope
WACO targets manipulation policies fine-tuned on a few dozen demonstrations and addresses camera and robot-base pose shifts, evaluated in simulation and on a UR5e robot.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.