Reproduce the bimanual valve study¶

All three modes score 30/30 on the fixed paired test cohort. The commands below require the completed study's frozen-sources.json. Do not generate a replacement manifest to bypass a missing artifact or a source mismatch.

The simulation and renderer revision is 6cc568d08bc5db239cf07e180e95d1829fa46369. Revision 52be64c489304326ffee0a55993d2cbe0ceb3cab adds the documented coordinate-precision audit without changing any of those 749 frozen source, asset and environment files. Use the latter revision for simulation commands below. The original manifest was created after the simulation commit. A newer publication commit may add reports, paper sources and website material while retaining those exact runtime bytes. The protocol document itself is frozen: its development-era status paragraphs are historical provenance, not the current campaign status.

This is a scripted-control study on a custom twin-Oberon Rex, with a separate 506 mm passive valve. It does not replace the original RotateValve ACT/DP results or the native two-platform PressButton study. See the physical model and fixed protocol.

Obtain the recorded evidence¶

Run downloads from the current publication branch, before entering the pinned runtime: the older runtime commit predates the final artifact registry. The versioned bimanual-evidence artifact contains the original development, final and demo traces, sampled audits, coordinate counterfactual and the same-machine clean-install reproduction. Downloading it does not start a run:

python3 scripts/fetch_paper_artifacts.py fetch --artifact bimanual-evidence \
  --destination .maintenance/backups/bimanual-evidence-download \
  --cache .maintenance/backups/download-cache
mkdir .maintenance/backups/bimanual-evidence-view
python3 -m tarfile -e \
  .maintenance/backups/bimanual-evidence-download/bimanual-valve-v1-evidence.tar.gz \
  .maintenance/backups/bimanual-evidence-view

Use the archive's bimanual-valve-v1/manifest.json to inspect per-file hashes. campaign/summary.json reports the 90 powered test episodes; clean-reproduction/ contains three separate demo repeats and their motors-disabled controls. Failed preflights and original numerical diagnostics remain in diagnostics/. The bimanual-demo-free, bimanual-demo-support and bimanual-demo-two-hands artifact IDs download the three final silent pool films separately. The complete website bundle already includes those films.

Install and verify the runtime¶

From the current publication checkout, retain the manifest's absolute path and prepare the isolated runtime:

export WM_BIMANUAL_MANIFEST="$(pwd)/benchmarks/bimanual-valve-v1/frozen-sources.json"
python3 scripts/paper_release.py prepare --runtime bimanual \
  --directory .maintenance/checkouts/bimanual-v1
cd .maintenance/checkouts/bimanual-v1

Follow installation inside this checkout, skipping its clone step, and install the optional recording dependency:

uv pip install --python .venv/bin/python --no-deps -r research/media-requirements.txt
.venv/bin/python scripts/run_bimanual_study.py \
  --manifest "$WM_BIMANUAL_MANIFEST" \
  --mode two-hands --split final --output-dir artifacts/bimanual-replica/two-hands \
  --check-only

The launcher checks frozen source/asset hashes before running. --check-only prints the exact command and reset IDs; it does not evaluate success. Use the recorded Python/Isaac Lab/package versions and preserve authorized mesh files. Choose fresh output directories for every execution.

The pinned checkout does not itself contain the later benchmark manifest. The variable above points back to the publication copy. Alternatively set it to the extracted archive's bimanual-valve-v1/campaign/frozen-sources.json. This is an evaluation input, not a file to regenerate after inspecting test outcomes.

Repeat the three-condition benchmark¶

Run one physics process at a time. Each mode uses the same 30 test reset IDs 92000–92029, a 180-second physical horizon, 240 Hz physics and 30 Hz commands. An extra motors-disabled control is simulated but excluded from the denominator.

set -e
for mode in free support two-hands; do
  OMP_NUM_THREADS=4 .venv/bin/python scripts/run_bimanual_study.py \
    --manifest "$WM_BIMANUAL_MANIFEST" \
    --mode "$mode" --split final --output-dir "artifacts/bimanual-replica/$mode"
  .venv/bin/python scripts/analyze_bimanual_trace.py "artifacts/bimanual-replica/$mode"
  .venv/bin/python scripts/audit_bimanual_trace.py "artifacts/bimanual-replica/$mode"
  .venv/bin/python scripts/audit_bimanual_coordinate_precision.py \
    "artifacts/bimanual-replica/$mode" \
    --output "artifacts/bimanual-replica/$mode-coordinates"
done
.venv/bin/python scripts/report_bimanual_coordinate_study.py \
  --runs artifacts/bimanual-replica/free artifacts/bimanual-replica/support \
         artifacts/bimanual-replica/two-hands \
  --coordinate-audits artifacts/bimanual-replica/free-coordinates \
         artifacts/bimanual-replica/support-coordinates \
         artifacts/bimanual-replica/two-hands-coordinates \
  --split final --output artifacts/bimanual-replica/summary.json
.venv/bin/python scripts/plot_bimanual_study.py artifacts/bimanual-replica/summary.json \
  --output-dir artifacts/bimanual-replica/plots

free parks the left arm; support grasps the fixed rail with it; two-hands turns the wheel with both hands. The first two modes retain identical fixture geometry. In the two-hand extension the unused rail is moved aside. The base is free floating in all three conditions.

Success remains sampled signed shaft rotation of at least 170 degrees. Contact, tracking and actuator-load diagnostics do not silently add success criteria. The separate CAD audit checks sampled robot self-collision, native joint limits, motor force limits and forward kinematics. Read audit reports as well as the endpoint result. Contact work estimates use 30 Hz samples; they are not actuator energy measurements. Orientation diagnostics use float64 quaternion geometry to retain small rotations that float32 inverse-cosine diagnostics can round to zero. The supplemental FK audit separates float32 world-coordinate rounding from kinematic disagreement; all three residual checks retain the 100 µm bound. The original absolute-world discrepancies remain recorded. This read-only diagnostic amendment does not rerun or replace physical episodes.

Record the separate pool demonstration¶

Demo reset 91003 is separate from the test cohort. Record its physical motion first, then render the measured states. This avoids concurrent camera/dynamics workloads without modifying the simulated trajectory.

OMP_NUM_THREADS=4 .venv/bin/python scripts/run_bimanual_study.py \
  --manifest "$WM_BIMANUAL_MANIFEST" \
  --mode two-hands --split demo --physics-only \
  --output-dir artifacts/bimanual-demo/two-hands
.venv/bin/python scripts/analyze_bimanual_trace.py artifacts/bimanual-demo/two-hands
.venv/bin/python scripts/audit_bimanual_trace.py artifacts/bimanual-demo/two-hands
.venv/bin/python scripts/audit_bimanual_coordinate_precision.py \
  artifacts/bimanual-demo/two-hands --output artifacts/bimanual-demo/two-hands-coordinates
OMP_NUM_THREADS=4 .venv/bin/python scripts/render_bimanual_trace.py \
  --run artifacts/bimanual-demo/two-hands \
  --output-dir artifacts/bimanual-demo/two-hands/pool-render

Use support and free in fresh directories for the other two films. Output is 180 seconds at 1280×720, 30 fps, real time, silent and without text overlays. observer.json records trace/asset hashes, tool-pose agreement and clearance from the pool floor, walls and water surface. --snapshots-only produces a framing preview marked incomplete; it cannot substitute for a final movie.

The existing pool and textured wall are presentation geometry. Rendering does not add boundary hydrodynamics, rerun physics or change success. Rotor animation follows recorded motor forces with slowed display rotation, not physical RPM.

Interpret a reproduction¶

The original development comparison checks all three modes at 240 and 480 Hz on reset IDs 91000–91002 before opening the final cohort. Repeating a published cohort is reproduction, not a new independent test for a subsequently tuned controller. Preserve failed runs and use separate output paths.

A clean environment on the same workstation verifies installation and recorded execution to its stated scope. It does not establish independent-machine or hardware transfer. The reproduction record identifies the tested commit, reset IDs, full horizon and numerical discrepancies; package versions are fixed by the original source manifest.