facebookresearch/theseus

Embeds differentiable second-order optimizers (Gauss-Newton, Levenberg-Marquardt, Trust Region) and sparse linear solvers directly into PyTorch, enabling end-to-end gradient flow through optimization layers. Supports batched GPU computation with multiple backward modes (implicit, truncated, DLM) and includes Lie group representations and robot kinematics for robotics and vision applications. Designed for hybrid architectures that combine neural networks with domain-specific differentiable models as inductive priors.

2,008 stars. No commits in the last 6 months.