ADM-Fusion adapts sensor weighting for robust robot ego-motion
Deep learning fusion handles sensor failures with adaptive weighting and cross-task attention.
ADM-Fusion tackles a critical challenge in autonomous robotics: maintaining accurate ego-motion estimation when individual sensors fail or degrade. Traditional fusion methods rely on fixed weighting, but ADM-Fusion introduces an adaptive sensor mixture-of-experts framework with content-aware routing that rebalances sensor contributions in real time based on environmental conditions and degradation levels. The system also separates translation and rotation estimation into distinct branches, connected via a cross-task attention mechanism that preserves task-specific specialization while enabling information sharing.
Training on the CARLA-LOC simulated dataset followed by fine-tuning on real-world KITTI data demonstrates effective simulation-to-real transfer. Experiments show ADM-Fusion maintains competitive performance against existing methods even under severe sensor degradation, such as blocked cameras or noisy IMU readings. This approach is particularly valuable for autonomous vehicles and robots operating in diverse, unpredictable environments where sensor reliability can fluctuate rapidly.
- Uses content-aware routing in a mixture-of-experts to dynamically assign weights to sensor inputs
- Separates translation and rotation estimation with cross-task attention for better specialization
- Trained on simulated CARLA-LOC and fine-tuned on real-world KITTI with robust performance under sensor degradation
Why It Matters
Enables autonomous vehicles and robots to maintain accurate localization even when sensors degrade in harsh environments.