Problem formulation
Intent aliasing describes opposing actions under nearly identical observations.
Language-conditioned control · Robotic endoscopy
EndoLIFT
Language-Disambiguated Latent-Conditioned Rectified Flow for Bidirectional Endoscopic Control
The image describes the anatomy. Language specifies the direction.
1The Chinese University of Hong Kong · 2The Sixth Affiliated Hospital, Sun Yat-sen University · Preprint arXiv:2608.20478
arXiv 2026
+11.1Points in direction accuracy
−83%Wrong-direction advance
10 / 10Ex-vivo trials completed
01 / Overview
Intent aliasing
The same view can require the opposite action.
Endoscopy is bidirectional: the instrument advances to reach the anatomy and later withdraws for inspection. A physiological alert or verified voice instruction may request an earlier reversal, before the visual scene has changed.
Pixels alone cannot identify which procedural phase is active. EndoLIFT supplies the missing variable through language and uses one shared policy for forward navigation and urgent retraction.
Language interface
Canonical instructions select the active axial mode.
Trajectory latent
A latent-conditioned flow expert generates continuous action chunks.
02 / Method
Architecture
Select the intent, condition the trajectory, then act.
A PaliGemma 2 backbone encodes the image and instruction. The previous action, visual-language prefix, flow time, and a 32-dimensional variational trajectory latent condition an eight-block Transformer action expert.
The expert generates a 32-step action chunk for longitudinal motion, up–down bending, and left–right bending. During deployment, a change in instruction invalidates the active chunk and immediately triggers replanning.
03 / Project video
Full presentation
Problem, model, and experiments in six minutes.
The complete presentation covers intent aliasing, language conditioning, latent rectified flow, phantom transfer, and ex-vivo evaluation.
04 / Results
Controlled evaluation
Language selects direction; the latent improves control.
Same-observation instruction swaps show that language changes the axial mode independently of latent conditioning. Relative to the matched model without the trajectory latent, EndoLIFT improves direction accuracy by 11.1 points and reduces wrong-direction advance by 83%.
Across 44 held-out language variants, intent-following accuracy is 82.8%. Closed-loop evaluation improves success by 30 points in the seen colon phantom and unseen lung and stomach phantoms, followed by 10 successful ex-vivo porcine-trachea trials.
05 / Citation
Publication
EndoLIFT.
Preprint, 2026.
@article{ng2026endolift,
title={EndoLIFT: Language-Disambiguated
Latent-Conditioned Rectified Flow for
Bidirectional Endoscopic Control},
journal={arXiv preprint arXiv:2608.20478},
year={2026}
}