CredibleDFGO: Differentiable GNSS Factor Graph with Credibility Supervision
A new differentiable factor graph optimization framework, CredibleDFGO (CDFGO), addresses unreliable covariance estimates in GNSS positioning for urban navigation. Existing DFGO methods learn measurement weighting but use position-only objectives, leading to inaccurate reported covariance. CDFGO makes covariance credibility an explicit training target. Its Weighting Generation Network (WGN) predicts per-satellite reliability weights, which are mapped by a differentiable Gauss-Newton solver to a position estimate and posterior covariance. Proper scoring rules supervise the East-North predictive distribution end-to-end, studying negative log-likelihood (NLL) and Energy Score.
Key facts
- CredibleDFGO (CDFGO) is a differentiable GNSS factor graph framework.
- It targets covariance credibility explicitly in training.
- Weighting Generation Network (WGN) predicts per-satellite reliability weights.
- Differentiable Gauss-Newton solver maps weights to position estimate and posterior covariance.
- Proper scoring rules supervise East-North predictive distribution end-to-end.
- Studies negative log-likelihood (NLL) and Energy Score.
- Addresses unreliable covariance from GNSS solvers in urban canyons.
- Existing DFGO methods use position-only objectives, leading to inaccurate covariance.
Entities
—