Problem

Multi-robot collaboration in scenarios like search-and-rescue faces significant challenges. Robots often have limited onboard resources for sensing, computing, and communication, hindering real-time data collection and efficient processing. Dynamic and complex environments make accurate path planning difficult and can degrade wireless communication quality. Existing server-centric centralized architectures introduce high latency, preventing timely decision updates crucial for time-critical missions. While edge computing helps by bringing resources closer, and Large AI Models (LAMs) offer strong capabilities for task analysis and decision support, deploying LAMs alone at the edge can lead to excessive inference delays and costly updates in rapidly changing environments, making it difficult to guarantee real-time decision output.

Method

The proposed LSAI (Large Small AI model codesign) framework addresses these issues by deeply integrating large and small AI models. Robots collect heterogeneous information using multimodal sensors and train local Small AI (SAI) models with their onboard resources. The lightweight parameters of these SAI models are then uploaded to an edge server. At the edge, an attention-based model aggregation algorithm fuses these parameters to construct a global Large AI (LAI) model, which more accurately estimates complex environments and predicts changes. The LAI model also generates optimal sensing path planning decisions by considering robot energy, location, and velocity. Simultaneously, an adaptive model splitting and update algorithm extracts suitable parameters from the LAI to update the local SAI models, enabling robots to analyze local environmental variations and refine their sensing paths online for energy efficiency and collision avoidance.

Results

Simulations were conducted in a 3km x 3km robot-based search and rescue scenario using Gazebo, with up to 60 robots. The LSAI framework demonstrated superior performance compared to traditional centralized large AI model implementations and distributed small AI model methods. LSAI consistently achieved higher sensing accuracy, with performance increasing monotonically with the number of robots and showing a wider gap over benchmarks as robot count grew. It attained up to 20.4% higher sensing accuracy than traditional solutions. The path planning efficiency of LSAI was consistently the highest across all settings, exhibiting the smallest performance degradation and an average efficiency of approximately 0.936, significantly outperforming the distributed (0.822) and centralized (0.746) schemes. Furthermore, LSAI consistently achieved the lowest system implementation latency, reducing sensing cooperation latency by an average of 17.9% and showing superior efficiency and scalability as the number of robots increased, with an average latency of approximately 13 minutes compared to 20.2 minutes for distributed and 24.6 minutes for centralized approaches.

Implications

The LSAI framework provides a robust architectural foundation for future multi-robot intelligent applications by effectively integrating terminal-edge computing resources and enhancing collaborative sensing in complex environments. This codesign approach allows for more adaptive, efficient, and trustworthy robot cooperation through dynamic role allocation, where large models handle high-level reasoning and small models perform fast local inference. Potential applications extend to intelligent healthcare and wearable monitoring, enabling early warning systems and personalized assistance by leveraging edge devices for continuous local monitoring and escalating complex cases to more powerful models. However, limitations remain in ensuring perfect environment sensing due to inherent robot capabilities and challenges in dynamic neighbor selection given limited computing resources.