Skip to content

plan.md

Latest commit

 

History

History
139 lines (98 loc) · 4.71 KB

plan.md

File metadata and controls

139 lines (98 loc) · 4.71 KB
Raw

CudaRobotics Plan / Handoff

Last updated: 2026-04-22 JST

This is the short handoff for the next coding session. The repo is currently on master, synced with origin/master, and the working tree was clean before this file update.

Current State

  • Main branch: master
  • Latest merged work:
    • PR #1: reproducible benchmark suite runner
    • PR #2: point-cloud CLI demo workflow
  • Recent merge commit: b4f1bb0 from PR #2
  • CI status for PR #2: passed
  • Deferred branch: feat/repro-report exists but was intentionally not merged because the user wanted to move to higher-value work.

What Is Done

Reproducible Benchmark Suite

The repo now has a reproducible benchmark runner path that can be used to keep experiments repeatable. This is useful infrastructure, but it is not the most interesting product-facing value right now.

Point-Cloud CLI Demo

bin/benchmark_pointcloud is now usable as a small CLI demo, not only as a fixed benchmark.

Supported workflow:

./bin/benchmark_pointcloud --quick
./bin/benchmark_pointcloud --xyz examples/pointcloud/sample_room.xyz --input-only --op voxel --leaf-size 0.8 --out build/sample_room_voxel.ply
./bin/benchmark_pointcloud --xyz examples/pointcloud/sample_room.xyz --input-only --op ransac --plane-threshold 0.05 --out build/sample_room_plane.ply
./bin/benchmark_pointcloud --xyz examples/pointcloud/sample_room.xyz --input-only --op normals --k 12 --out build/sample_room_normals.ply

Useful options:

  • --quick
  • --input-only
  • --op voxel|statistical|normals|ransac|gicp|all
  • --out PATH
  • --leaf-size, --k, --std-mul, --plane-threshold, --ransac-iters, --seed

Bundled sample:

  • examples/pointcloud/sample_room.xyz

Outputs:

  • voxel/statistical/ransac: .ply or .xyz point rows
  • normals: .ply with normal properties, or .xyz rows as x y z nx ny nz

Validation Already Run

cmake --build build --target benchmark_pointcloud -j$(nproc)
./bin/benchmark_pointcloud --help
./bin/benchmark_pointcloud --quick --op voxel
./bin/benchmark_pointcloud --xyz examples/pointcloud/sample_room.xyz --input-only --op voxel --leaf-size 0.8 --out build/sample_room_voxel.ply
./bin/benchmark_pointcloud --xyz examples/pointcloud/sample_room.xyz --input-only --op ransac --plane-threshold 0.05 --ransac-iters 256 --out build/sample_room_plane.ply
./bin/benchmark_pointcloud --xyz examples/pointcloud/sample_room.xyz --input-only --op normals --k 12 --out build/sample_room_normals.ply
./bin/benchmark_pointcloud --xyz examples/pointcloud/sample_room.xyz --input-only --op normals --k 12 --out build/sample_room_normals.xyz
ctest --test-dir build -L 'python|cpu' --output-on-failure
git diff --check

Next High-Value Directions

1. Make The Point-Cloud Demo More Visual

Best next step if the goal is OSS value:

  • add a small examples/pointcloud/README.md
  • document before/after outputs for voxel, RANSAC plane, and normals
  • add a script that runs the sample demo commands and writes all artifacts into build/pointcloud_demo/
  • optionally add screenshots or a simple viewer command path if the repo already has a lightweight dependency available

Why: the project becomes easier to understand in 60 seconds. This is more visible than another internal benchmark report.

2. Add A Real Registration Demo

The current external-input path does not write a GICP result because GICP needs a source and target cloud. A valuable follow-up would add:

  • --source PATH
  • --target PATH
  • --op gicp
  • optional synthetic transform mode for the bundled sample cloud
  • output of transformed source cloud and final transform matrix

Why: registration is one of the most recognizable point-cloud robotics workflows.

3. Keep Diff-MPPI As A Research Track

Diff-MPPI remains the strongest research story in the repo. Return to it when the goal is paper value:

  • final paper wording
  • exact-time table cleanup
  • standardized benchmark expansion
  • stronger MuJoCo or hardware-facing task

Why: this is likely the best academic/research value, but it is heavier than a quick OSS demo improvement.

4. Do Not Resume feat/repro-report By Default

The repro report renderer is useful, but the user explicitly called it less interesting. Treat it as parked unless the next goal is experiment governance or release engineering.

Recommended Next Session

Start from master, pull, then build the next point-cloud user-facing feature:

git pull --ff-only
cmake --build build --target benchmark_pointcloud -j$(nproc)

Recommended branch name:

git switch -c feat/pointcloud-registration-demo

Recommended target:

Build a source/target GICP demo with generated transform output and a short example README. That is the clearest next value after the point-cloud CLI merge.