Co-GLANCE: Uncertainty-Aware Active Perception for Heterogeneous Robot Teaming

Researchers introduce Co-GLANCE, an onboard AI system for multi-robot teams that detects and resolves perceptual uncertainty in unstructured environments without cloud computing. By distilling vision-language model capabilities into an efficient local model with statistical uncertainty guarantees, the system achieves 25-36% accuracy improvements over cloud-based approaches while reducing inference latency by 350x.

Co-GLANCE addresses a fundamental challenge in autonomous robotics: enabling heterogeneous teams to collaborate effectively despite incomplete scene understanding caused by occlusions and viewpoint limitations. Traditional approaches either rely on centralized cloud-based vision-language models, introducing latency and computational overhead, or fail to quantify confidence in their predictions. This research bridges that gap by embedding sophisticated semantic reasoning directly into edge-deployed systems.

The innovation combines three core technical contributions: knowledge distillation from vision-language models to reduce computational requirements, conformal prediction for statistically valid uncertainty quantification, and a robot allocation mechanism that optimally assigns team members to resolve ambiguous regions. The 350x latency reduction and significant accuracy improvements suggest the approach scales practical robotic deployments beyond controlled laboratory settings.

For the robotics and autonomous systems industry, this work demonstrates that edge intelligence doesn't require sacrificing reasoning quality. The explicit uncertainty quantification is particularly valuable—it allows systems to acknowledge limitations rather than making confident wrong decisions, essential for safety-critical applications in exploration, search-and-rescue, and industrial inspection.

The release of an air-ground dataset and open-source code accelerates adoption of these techniques across the research community. Future development likely focuses on extending these methods to larger, more diverse robot teams and increasingly complex environmental conditions. The approach's emphasis on statistical guarantees and explainable confidence scores positions it well for regulatory compliance in emerging autonomous systems frameworks.

• →Co-GLANCE achieves 350x latency reduction compared to cloud-based vision-language models while improving accuracy by 25-36%.
• →The system uses conformal prediction to provide statistically valid uncertainty estimates that directly trigger active perception in robot teams.
• →Knowledge distillation enables complex semantic reasoning on resource-constrained onboard hardware without cloud dependency.
• →Calibrated uncertainty quantification allows autonomous systems to reliably acknowledge limitations rather than make overconfident incorrect decisions.
• →Open-source release of code, videos, and datasets accelerates research adoption in heterogeneous robot collaboration.