CoPeD

Zhou 2024 · Paper

One-line summary — CoPeD is a comprehensive real-world multi-robot collaborative perception dataset built around air-ground robot teams, created to fill the evaluation gap that had kept collaborative perception research stuck in simulation.

Problem

In the past decade single-robot perception made significant advances, but multi-robot collaborative perception remained largely unexplored — and one major hurdle was the lack of real-world datasets. Collaborative perception requires fusing compressed, intermittent, limited, heterogeneous, and asynchronous environmental information across robots, under sensor noise, occlusions, and outright sensor failures. Existing multi-robot datasets were predominantly designed for SLAM (geometric mapping with limited cross-robot view overlap), while collaborative-perception datasets were V2V autonomous-driving benchmarks, mostly simulated and missing aerial viewpoints. CoPeD was built specifically to close this gap.

Method & architecture

As a dataset paper, CoPeD’s “method” is its collection, calibration, and annotation pipeline (it contains no core equations):

Results

The paper’s evaluation is qualitative — its contribution is infrastructural. It demonstrates the dataset on collaborative perception tasks: a graph-neural-network system where two aerial robots and one ground robot share feature maps performs multi-robot monocular depth estimation and semantic segmentation, recovering predictions of far-away objects (trees in depth, houses in segmentation) and staying robust when one robot’s image sensor is corrupted, compared to a single-agent baseline. The zero-shot semantic/depth annotations are shown across indoor and outdoor sequences; see the paper for the full sequence tables and sensor specifications. Data and tools are released at arplaboratory/CoPeD.

Why it matters for SLAM

Collaborative SLAM papers have historically been evaluated on single-robot datasets artificially split among virtual agents, which hides the hard parts of real deployments: asynchronous sensing, heterogeneous platforms, and genuinely different viewpoints of the same place. CoPeD gives C-SLAM systems such as Kimera-Multi and Swarm-SLAM a realistic air-ground benchmark, much as EuRoC did for single-robot VIO — and its AprilTag-based cross-robot ground truth directly supports evaluating inter-robot loop closure and map merging.