Search papers, labs, and topics across Lattice.
This paper introduces ADM-Fusion, an adaptive deep learning framework for multi-sensor fusion aimed at robust ego-motion estimation in challenging environments. By employing a mixture-of-experts approach with content-aware routing, ADM-Fusion dynamically adjusts sensor weights in real-time, effectively integrating complementary cues from various modalities. Experimental results on both simulated and real-world datasets reveal that ADM-Fusion outperforms existing methods, demonstrating resilience to sensor degradation while maintaining high accuracy in ego-motion tasks.
ADM-Fusion achieves robust ego-motion estimation by dynamically adapting sensor weights in real-time, outperforming traditional fixed-weight methods under degraded conditions.
Robust multi-sensor fusion is essential for reliable autonomy in diverse and degraded environments, where sensor reliability can fluctuate rapidly. Because different modalities fail in distinct ways, effective fusion should adaptively balance complementary cues rather than rely on fixed weighting. This adaptability is particularly important for ego-motion estimation, since accurate updates depend on the consistent integration of complementary sensor information. We propose ADM-Fusion, an end-to-end deep learning based multi-sensor fusion method designed to adapt to environmental changes and sensor degradation. ADM-Fusion employs an adaptive sensor mixture-of-experts framework with content-aware routing to dynamically assign weights to sensor inputs in real time. The system further incorporates separate translation and rotation branches, coupled through a cross-task attention mechanism to preserve task-specific specialization while enabling information sharing. ADM-Fusion is trained on the CARLA-LOC simulated dataset and subsequently fine-tuned on KITTI real-world data, demonstrating effective simulation-to-real transfer. Experiments show that ADM-Fusion remains robust under degraded conditions while maintaining competitive performance against existing methods.
