CERPE: Communication-Efficient Relative Pose Estimation

Overview

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 large-scale pre-trained vision 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.

TL;DR: Screen overlap cheaply, request images selectively, and keep relative pose continuous when the shared view disappears.

Method

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

Overview

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.

Click the figure to enlarge it.

Results

We evaluate CERPE in CARLA simulation, real-world connected autonomous driving, and indoor physical robot teams. Click a figure to enlarge it.

Qualitative results

Error analysis

Error vs. overlap and viewing angle

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.

Quantitative results

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

Swipe sideways to see all columns.

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

Online deployment

Real-time relative pose estimation on robot teams, running 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

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

@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}
}

Contact: [email protected], [email protected]