theseus

Why we included this project

Robotics and vision systems often need to combine learned components with well-understood geometric constraints, and Theseus makes that combination trainable end to end. You define an optimization objective that mixes both kinds of terms, then let the library solve it with Gauss-Newton, Levenberg-Marquardt, trust region, or dogleg while gradients still flow back through the solver. That means a robot can learn from data while staying faithful to kinematic or physical priors. The library also offers sparse linear solvers, batching, GPU acceleration, and several backward modes, so you can balance training speed against approximation quality. It's a mature, principled tool that.