---

name: gradslam description: "Routes GradSLAM package workflows for RGB-D structures, tensor geometry, odometry and SLAM, dataset adapters, and CfgNode configuration on a CPU-first installation." disable-model-invocation: true metadata: disco-role: operating license: MIT

GradSLAM operating guide

Use this repo skill when a Researcher needs the public gradslam Python package for dense RGB-D geometry, RGB-D/point-cloud structures, odometry, PointFusion or ICPSLAM, TUM/ICL/ScanNet loading, or CfgNode configuration. This is operating context for a later Researcher, not a benchmark recipe and not a replacement for the package's external datasets.

Start here

1. Create an isolated environment, then install the public release with python -m pip install gradslam. When the exact source baseline is required, install the pinned public Git revision instead: python -m pip install 'git+https://github.com/gradslam/gradslam.git@44470eee4484aaad703d7bf962a8f42496867a9d'.
2. Run python -m pip check, then resolve and run the bundled environment diagnostic scripts/check_environment.py. Start with --help; the script reports import and backend facts without installing or changing anything.
3. Read compatibility before selecting a dependency combination. CPU is the required verified path; successful CPU execution does not certify CUDA.
4. Record the input tensor contract, device, image layout, depth units, pose availability, and external-data boundary before calling package APIs.
5. Route to exactly one focused sub-skill, then use its API/reference page and its deterministic helper where one exists.

Route by user intent

• RGB-D tensors, vertex/normal maps, point-cloud conversion, ragged clouds, or safe display adapters: structures.
• Projection, unprojection, intrinsics, pixel grids, rigid transforms, quaternions, or SE(3): geometry.
• Known-pose, ICP, GradICP, PointFusion, ICPSLAM, map fusion, or odometry diagnostics: odometry-slam.
• TUM, ICL-NUIM, ScanNet paths, association metadata, preprocessing, labels, or DataLoader handoff: datasets.
• Nested config trees, YAML/Python files, merge precedence, overrides, freezing, or migration diagnostics: configuration.

Cross-skill routing is intentional: datasets produce the RGB-D batch, structures validate its layout, geometry supplies transforms, and odometry/SLAM consumes the resulting objects. Do not skip the upstream shape and units check.

Package boundaries

• The package imports Open3D during import gradslam; a missing or incompatible optional-looking dependency can therefore be an import blocker.
• The historical package metadata permits older Torch/dependency combinations, but exact compatibility is release- and extension-dependent. Follow the compatibility and troubleshooting references instead of blindly upgrading one binary package.
• External TUM, ICL-NUIM, and ScanNet data are not bundled. Layout checks are read-only preflight, not proof that images, poses, or timestamp association are valid.
• The bundled smoke programs use tiny deterministic in-memory CPU fixtures. They do not validate an external sequence, a full benchmark, CUDA behavior, a GUI viewer, or notebook execution.
• Keep colors, depths, intrinsics, poses, and map tensors on a deliberate device. Keep names, transforms, timestamps, semantic labels, and config metadata outside RGBDImages unless an API explicitly accepts them.

Self-contained references

• API surface — package exports and direct-module entry points.
• Compatibility — supported claims, dependency caveats, and CPU/CUDA boundaries.
• Troubleshooting — install/import, extension, device, data, and cross-workflow failures.
• Provenance — source version and evidence baseline.
• Environment checker — no-mutation import and backend diagnostic.

Safe operating sequence

Prefer a short, observable progression: import check → tiny tensor or config fixture → focused sub-skill smoke → one small caller-owned data item → only then longer SLAM or dataset work. Stop when a required invariant fails. Do not invent poses, silently rescale depth, download data, open a viewer, or claim an accelerator path from a CPU-only result.