ALIVE-LIO: Degeneracy-Aware Learning of Inertial Velocity for Enhancing ESKF-Based LiDAR-Inertial Odometry
A new hybrid AI system reduces robot pose drift by selectively using neural networks only when sensors fail.
A research team from KAIST and Seoul National University has introduced ALIVE-LIO, a novel framework designed to solve a persistent problem in robotics: accurate navigation in 'degenerate' environments. Traditional LiDAR-inertial odometry (LIO) systems, which fuse laser scans and inertial data to track a robot's movement, often fail in feature-poor spaces like long hallways, tunnels, or near single walls. This is because LiDAR loses observability in certain directions, causing the robot's estimated position to drift uncontrollably. ALIVE-LIO's key innovation is its degeneracy-aware design, which strategically employs a deep neural network to predict inertial velocity only when the system detects these problematic conditions.
Instead of replacing the proven probabilistic framework of an Error-State Kalman Filter (ESKF), ALIVE-LIO enhances it. The neural network's velocity predictions are selectively fused into the ESKF's state update, specifically compensating for the directions where LiDAR data is unreliable. This hybrid approach maintains the filter's consistency and structure while gaining the adaptive power of machine learning. In rigorous testing on public datasets and newly collected data, ALIVE-LIO delivered the most competitive results in 22 out of 32 sequences, dramatically cutting pose drift. The team has committed to making the implementation publicly available, paving the way for more robust autonomous vehicles, drones, and mobile robots that must operate reliably in challenging, real-world geometries.
- Hybrid AI/Classical Filter: Integrates a deep neural network for velocity prediction into a standard Error-State Kalman Filter (ESKF), enhancing rather than replacing classical methods.
- Degeneracy-Active Design: The AI component is only activated upon detecting sensor 'degeneracy' (e.g., in long corridors), making it efficient and targeted.
- Superior Performance: Outperformed other methods, achieving the best results in 22 out of 32 test sequences on challenging datasets.
Why It Matters
Enables robots and autonomous systems to navigate reliably in real-world, feature-poor environments where current systems fail, critical for practical deployment.