KAIST's delay-aware RL cuts drone tracking errors by 27%

KAIST researchers have developed a delay-aware, uncertainty-driven multi-agent reinforcement learning framework for active triangulation of aerial targets in Counter-UAS applications. Traditional active triangulation methods assume instantaneous state feedback, ignoring cumulative latency from detection, communication, and decision propagation. The team formalized the problem as a Dec-POMDP (Decentralized Partially Observable Markov Decision Process) with Age-of-Information (AoI) augmented observations, enabling delay-aware coordination among mobile observers with controllable cameras.

In experiments using MAPPO across 4096 parallel environments, the framework achieved 0.547 ± 0.217 m RMSE with 78.1% triangulation validity. AoI improved triangulation validity by 10.6 percentage points. Perception-consistent rewards outperformed privileged clean-state rewards (0.547 m vs. 0.633 m RMSE, 27% fewer track losses). Multi-source analytical covariance propagation incorporating pixel, pose, gimbal, and intrinsics uncertainties proved critical—restricting to angular noise alone caused 2.8-fold RMSE degradation. The paper (arXiv:2607.05957) also showed MLP policies achieve near-zero validity (0.7%), confirming that recurrent memory (e.g., LSTM) is essential for delay compensation. Accepted to IROS 2026.