PyPose is a robotics-focused library built on PyTorch that enables developers to integrate geometric and learning-based methods for tasks such as SLAM, model predictive control, and motion planning. It provides deep PyTorch integration, allowing users to leverage familiar syntax while applying bundle adjustment, factor graph optimization, and neural feature extraction within a single framework.
Funding
Funding not disclosed
Founders
Product
Problem
Robotic developers often need to implement complex perception and control pipelines that combine geometric algorithms with deep learning, but existing tools require switching between disparate libraries and learning new APIs, leading to increased development time and integration errors.
Solution
PyPose provides a unified robotics software library built on PyTorch, allowing developers to write SLAM, model‑predictive control, and motion‑planning code using familiar PyTorch syntax. The library offers deep integration with PyTorch’s autograd and tensor operations, enabling seamless blending of geometry‑based methods such as bundle adjustment and factor‑graph optimization with learning‑driven components like feature extraction and loop‑closure detection. By consolidating these capabilities in a single framework, PyPose reduces the overhead of managing multiple toolchains and accelerates the development of robust, learning‑enhanced robotic systems.
Target Audience
Primary users are robotics researchers, engineers, and developers building perception, control, and planning systems who prefer a PyTorch‑centric workflow.
Features
- Native PyTorch integration with full support for autograd, tensor operations, and GPU acceleration
- Modules for visual‑inertial, stereo, multi‑view, and LiDAR SLAM pipelines
- Model‑predictive control (MPC) and motion‑planning components compatible with PyTorch models
- Hybrid geometric‑learning workflows that combine bundle adjustment, factor‑graph optimization, and neural network‑based feature extraction
- Extensible API that allows custom learning modules to be inserted into traditional robotics algorithms