MATT-Diff: Multimodal Active Target Tracking by Diffusion Policy

TL;DR

MATT-Diff tackles active multi-target tracking under unknown target counts and dynamics by learning a multimodal diffusion policy conditioned on egocentric map tokens and Gaussian target beliefs. It combines three expert-planner demonstrations to cover exploration, tracking, and reacquisition, and uses a vision-transformer map encoder with a target-attention module to handle variable numbers of targets. Trained as a conditional diffusion model, it generates multi-modal action sequences that implicitly switch among exploration, tracking, and reacquisition, yielding lower target-uncertainty metrics (entropy and NLL) while maintaining competitive RMSE. The work demonstrates the practical potential of diffusion policies for complex robotic tracking tasks with limited sensing, suggesting strong generalization to dynamic, uncertain environments and providing a foundation for further RL-based fine-tuning and advanced target estimation methods.

Abstract

This paper proposes MATT-Diff: Multi-Modal Active Target Tracking by Diffusion Policy, a control policy that captures multiple behavioral modes - exploration, dedicated tracking, and target reacquisition - for active multi-target tracking. The policy enables agent control without prior knowledge of target numbers, states, or dynamics. Effective target tracking demands balancing exploration for undetected or lost targets with following the motion of detected but uncertain ones. We generate a demonstration dataset from three expert planners including frontier-based exploration, an uncertainty-based hybrid planner switching between frontier-based exploration and RRT* tracking based on target uncertainty, and a time-based hybrid planner switching between exploration and tracking based on target detection time. We design a control policy utilizing a vision transformer for egocentric map tokenization and an attention mechanism to integrate variable target estimates represented by Gaussian densities. Trained as a diffusion model, the policy learns to generate multi-modal action sequences through a denoising process. Evaluations demonstrate MATT-Diff's superior tracking performance against expert and behavior cloning baselines across multiple target motions, empirically validating its advantages in target tracking.

Figures (3)

• Figure 1: Our MATT-Diff architecture consists of a map encoder and a target encoder, conditioned on the robot pose through coordinate transformations. The map encoder converts a local egocentric map into patch tokens via CNN and feeds them into a Performer transformer. The target encoder processes detected target beliefs with masking for undetected targets through self-attention to produce context-aware embeddings. A diffusion policy, then, performs a denoising process to generate multimodal action sequences.
• Figure 2: Temporal changes of entropy over a representative episode. The policies shown are frontier-based (blue), time-based (purple), and MATT-diff (green).
• Figure 3: Trajectories of MATT-Diff and the benchmark methods in one episode. MATT-Diff achieves good balance of exploration and detected target tracking.