Diffusion Policy with Bayesian Expert Selection for Active Multi-Target Tracking

A team of researchers has developed a novel AI framework that significantly improves how mobile robots track multiple moving targets simultaneously. The system, called Diffusion Policy with Bayesian Expert Selection, addresses a fundamental challenge in active multi-target tracking: balancing exploration for new targets with exploitation of already-tracked ones. Unlike existing diffusion policies that implicitly select strategies through denoising, this approach explicitly quantifies uncertainty about which expert strategy to execute, treating expert selection as an offline contextual bandit problem.

The core innovation is a Bayesian framework that enables pessimistic, uncertainty-aware strategy selection. A multi-head Variational Bayesian Last Layer (VBLL) model predicts the expected tracking performance of each expert strategy given the current belief state, providing both point estimates and predictive uncertainty. Following the pessimism principle for offline decision-making, a Lower Confidence Bound (LCB) criterion selects the expert whose worst-case predicted performance is best, avoiding overcommitment to experts with unreliable predictions. The selected expert then conditions a diffusion policy to generate corresponding action sequences.

In experiments on simulated indoor tracking scenarios, the approach demonstrated superior performance compared to both the base diffusion policy and standard gating methods, including Mixture-of-Experts selection and deterministic regression baselines. The framework's ability to explicitly quantify and leverage uncertainty in strategy selection represents a significant advancement over existing methods that rely on implicit strategy selection through the denoising process of diffusion models.