Pragmatic Communication in Multi-Agent Collaborative Perception

Area of Science:

• Computer Science
• Robotics
• Artificial Intelligence

Background:

• Collaborative perception enhances agent abilities through message exchange, but faces a trade-off between perception and communication costs.
• Existing methods transmit high-dimensional feature maps, leading to substantial communication overhead.
• Optimizing the perception-communication trade-off is crucial for efficient multi-agent systems.

Purpose of the Study:

• To develop a communication-efficient collaborative perception system that minimizes data transmission.
• To address the limitations of high communication costs in current collaborative perception approaches.
• To propose a pragmatic communication strategy focusing on task-critical information.

Main Methods:

• Formulated a mathematical optimization framework for the perception-communication trade-off.
• Proposed PragComm, a system featuring pragmatic message selection, representation, and collaborator selection.
• Implemented pragmatic message selection for spatially and temporally sparse feature vectors.
• Utilized a task-adaptive dictionary for pragmatic message representation, enabling communication via integer indices.
• Developed pragmatic collaborator selection to identify beneficial collaborators and prune unnecessary links.

Main Results:

• PragComm demonstrated superior performance in collaborative 3D object detection and tracking tasks.
• Achieved over 32.7K× lower communication volume compared to previous methods on the OPV2V dataset.
• The system adapts effectively to various communication conditions.
• Evaluated on real-world (V2V4Real) and simulation (OPV2V, V2X-SIM2.0) datasets.

Conclusions:

• PragComm offers a significant advancement in communication efficiency for collaborative perception.
• The pragmatic communication strategy effectively balances perception ability and communication costs.
• The proposed system shows strong potential for real-world multi-agent applications requiring efficient data exchange.