Skip to main content

Benchmarks and measured results

10 / 10four-part screw missions completed
687 smedian planning time
13 / 13failure episodes recovered
0native crashes

Four-part screw assembly​

The current end-to-end dataset contains ten seeded runs on Linux aarch64, Python 3.11, stock PyPI HPP 9.0.2 wheels, Pinocchio 4.1.0, and NumPy 2.4.6. Three runs executed concurrently. Each mission contains 19 blocks, 31 grasp/release phases, and five home moves.

MetricResultInterpretation
Missions completed10 / 10All seeds reached the terminal rack-driver block
Planning blocks140Excludes direct home moves
Blocks replanned from entry0 / 140Lookahead avoided irreversible commitments in this batch
Failure episodes recovered13 / 13All recovered through resume
Wall-clock time547.16–979.22 sIncludes planning, recovery, optimization, and logging
Median wall-clock time687.11 sRoughly 11.5 minutes
Native crashes0For this ten-run dataset

Per-seed time​

The time spread is dominated by path-verified lookahead: some seeds reject more clamp candidates before finding one that leaves both holes reachable. This benchmark measures planning in simulation against the configured collision and constraint model. It is not a physical assembly benchmark.

Component measurements​

MechanismBaselineResultScope
Sequence graph constructionO(N!) combinationsO(N) phase-local graphsAlgorithmic sequence scaling
8×7 graph generationover 20 min observed hangseconds after upstream fixGraph construction, not motion solve
Random sampling, 10k callsCORBA 1.2 sPyHPP 0.3 sHistorical microbenchmark
Configuration validation, 10kCORBA 0.8 sPyHPP 0.2 sHistorical microbenchmark
Path projection, 100 stepsCORBA 2.5 sPyHPP 0.8 sHistorical microbenchmark

The CORBA comparison comes from the predecessor engineering report and is retained as historical motivation. Current LongTAMP exposes PyHPP only; do not treat those numbers as a current cross-version benchmark.

Reproducibility​

The screw dataset is checked into the source repository as script/screw_assembly/results/pypi-wheel-batch-2026-09-26.json and identifies source commit 2117858. Reproduce it with the example’s batch runner and compare completion, recovery, replanning, wall-clock distribution, and native crashes—not only a single successful replay.