---

name: pypose description: "Use PyPose for differentiable robotics with PyTorch: Lie-group geometry, state estimation and control modules, dense or sparse nonlinear optimization, projection/spline utilities, and trajectory evaluation; route each task to the matching workflow and backend contract." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

PyPose

PyPose is a PyTorch-based library for differentiable robotics, Lie groups, and physics-based optimization. Use this skill when a task mentions PyPose, pypose, LieTensor, SO3/SE3, pose graphs, bundle adjustment, filters, MPC, IMU preintegration, ICP/PnP, splines, reprojection, APE/RPE, or second-order optimization.

Start here

1. Install the public package and a compatible PyTorch 2.x build: python -m pip install pypose (or install the repository package with its requirements/runtime.txt when working from a source checkout).
2. Check the package before writing a larger experiment: python scripts/check_pypose_env.py.
3. Decide the representation, module, objective, and device first. Keep all related tensors on one device and in a compatible floating dtype.
4. Read exactly one focused sub-skill below, then its nearest API/workflow references and troubleshooting guide. Use bundled smoke scripts for a small deterministic check before a dataset-scale run.

import torch
import pypose as pp

xi = pp.se3([0.1, -0.2, 0.05, 0.02, 0.03, -0.01])
pose = xi.Exp()
points = torch.randn(8, 3)
transformed = pose.Act(points)
assert transformed.shape == points.shape

Route by task

• lie-tensor — SO3/SE3/Sim3/RxSO3 groups and algebras; typed construction, Exp/Log, composition, inverse, point action, conversions, batching, gradients, and manifold parameters. Read sub-skills/lie-tensor/SKILL.md.
• robotics-modules — dynamics and systems, EKF/UKF/PF, LQR/MPC, IMUPreintegrator, EPnP, ICP, and GeodesicLoss. Read sub-skills/robotics-modules/SKILL.md.
• optimization — residual modules, GN/LM, solvers, damping strategies, robust kernels/correctors, schedulers, Jacobians, and BAE/CUDA sparse LM. Read sub-skills/optimization/SKILL.md.
• geometry-evaluation — camera projection and reprojection, point-set utilities, splines, downsampling, trajectory association and APE/RPE. Read sub-skills/geometry-evaluation/SKILL.md.

If a request crosses routes, normalize the data and transform conventions in lie-tensor first, then use the owning workflow: for example, SE(3)-parameter LM belongs to optimization, while an EKF over SE(3) states belongs to robotics-modules.

Device and optional backend policy

The base package requires PyTorch 2.x and runs on CPU or a compatible PyTorch accelerator. CUDA may accelerate ordinary operations, but do not call a CPU smoke proof of CUDA support. The sparse Jacobian/sparse LM route is different: it requires CUDA and the optional bae backend, with the verified compatible release bae==0.2.1; it has no CPU substitute. Read optimization/references/sparse-optimization.md before selecting sjac, psjac, PCG, or sparse=True.

Use the root diagnostic for an import/version/device report. Use sub-skills/optimization/scripts/sparse_lm_smoke.py with --check-only when sparse readiness is a hard requirement. If CUDA or BAE is absent, narrow the task to dense optimization or record the sparse capability as unverified; do not silently switch to CPU and claim equivalent coverage.

Shared references and maintenance

• Read references/troubleshooting.md for package import, PyTorch compatibility, device, optional dependency, and general shape failures.
• Read references/repo-provenance.md before deciding whether this skill matches a repository revision; refresh it when the package version, public entry points, or evidence baseline changes.
• references/repo-routing-metadata.json contains structured placement data for the managed repo-skill router; it is metadata, not a second usage guide.

The runtime files are self-contained. Original examples, tests, notebooks, data downloads, and plotting programs informed the references but are not runtime dependencies. Prefer the bundled helpers and small in-memory fixtures for verification. Do not copy private environment paths, activation commands, or checkout-specific imports into a Researcher task.