IROS 2026

CERPE

Communication-Efficient Relative Pose Estimation with Vision Foundation Models for Ephemeral Collaborative Perception

Qihang Li¹ Jo-Hao Huang¹ Jiewen Liu¹ Suyoung Kang² Hao Zhang² Peng Gao¹

¹ North Carolina State University ² University of Massachusetts Amherst

Paper Video

Overview

Relative pose through short, imperfect encounters

Relative pose estimation is fundamental to collaborative perception and coordination, yet robots often meet only briefly, communicate over limited bandwidth, and lose visual overlap through occlusion or restricted fields of view. Existing approaches commonly rely on global references, sustained overlap, or expensive visual transmission.

CERPE is a system-level framework that coordinates existing vision foundation models for metric ego-motion and inter-robot relative pose estimation. Robots continuously exchange fixed-size descriptors and local poses; descriptor similarity gates event-triggered raw-image requests. Direct estimates anchor the relative geometry when views overlap, while metrically scaled ego-motion propagates the latest anchor through non-overlapping intervals.

In one line: screen overlap cheaply, request pixels selectively, and preserve pose continuity when the shared view disappears.

Method

A conditional pipeline for communication and geometry

Descriptor screening is deliberately separated from pose inference, so raw RGB transmission occurs only after gate activation.

CERPE method overview showing descriptor-gated communication, direct relative pose estimation, and SE(3) propagation — Each direct cross-robot estimate becomes a pose anchor. When the gate is inactive, CERPE combines that anchor with both robots' metric ego-motion instead of requesting another remote image.

Screen

SALAD encodes each observation into a fixed-size descriptor. Descriptor similarity serves as an online proxy for candidate visual overlap.

Anchor

Gate activation requests the collaborator's image. VGGT estimates a direct pose, with metric scale recovered from aligned depth.

Propagate

Without a candidate overlap, the latest direct anchor is propagated in SE(3) using metric ego-motion from both robots.

Evaluation

Three settings, one continuous estimate

We evaluate CERPE in CARLA simulation, real-world connected autonomous driving, and indoor physical robot teams.

Method	CAD simulation			Real-world CAD			Physical robots
Method	Pos. m ↓	Rot. rad ↓	Succ. % ↑	Pos. m ↓	Rot. rad ↓	Succ. % ↑	Pos. m ↓	Rot. rad ↓	Succ. % ↑
SuperGlue	16.00^‡	0.895	77.5	7.62^‡	0.169	93.7	0.12^‡	0.165	98.4
NOPE	9.41	0.776	33.4	39.41	1.682	45.0	—	—	—
Co-VisNet	16.22	0.906	100	29.95	0.271	100	0.88	0.257	100
CERPE w/o EM	5.03 / 3.51^‡	0.128	100	10.27 / 5.84^‡	0.063	100	0.45 / 0.09^‡	0.096	100
CERPE	8.66 / 5.06^‡	0.089	100	8.26 / 2.94^‡	0.063	100	0.54 / 0.11^‡	0.086	100

^‡ Translation with ground-truth scale. CERPE rows report predicted-scale / GT-scale translation. Errors are reported over successful pairs.

CERPE qualitative results in CARLA — CARLA encounters transition between overlapping and non-overlapping views as vehicles cross an intersection. CERPE maintains a temporally stable relative trajectory through the transition.

Pose error across visual overlap ratios and relative viewing angles — CERPE remains comparatively stable across reduced visual overlap and large relative viewing angles. Propagation is a continuity mechanism, not a uniform accuracy boost: it can introduce scale and translation drift when direct overlap constraints are already strong.

Online deployment

Relative pose estimation on robot teams

Real-time estimation on a laptop.

CERPE Campus 01 demonstration.

Evaluation scope. These videos demonstrate online execution and system integration. Quantitative physical-robot accuracy is reported on the sampled motion-capture-covered scenarios in the paper.

Citation

BibTeX

Please cite the paper when using CERPE or its evaluation results.

Contact: qli39@ncsu.edu, pgao5@ncsu.edu

@inproceedings{li2026cerpe,
  title     = {Communication-Efficient Relative Pose Estimation with Vision Foundation Models for Ephemeral Collaborative Perception},
  author    = {Li, Qihang and Huang, Jo-Hao and Liu, Jiewen and Kang, Suyoung and Zhang, Hao and Gao, Peng},
  booktitle = {2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2026}
}