Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
The table of contents is too big for display.
Diff view
Diff view
  •  
  •  
  •  
54 changes: 44 additions & 10 deletions .github/workflows/ci.yml
Original file line number Diff line number Diff line change
Expand Up @@ -74,6 +74,7 @@ jobs:
dvidia-piper-grip --help
dvidia-piper-contact --help
dvidia-piper-recovery --help
dvidia-piper-reliability --help
dvidia-training-doctor --json
- name: Test footage, visual training and loopback studio
run: python -m unittest tests.test_footage_data tests.test_footage_train tests.test_footage_pipeline tests.test_footage_benchmark tests.test_training_studio tests.test_doctor tests.test_coverage tests.test_episode_tape tests.test_reaction tests.test_reaction_metrics tests.test_reaction_convergence tests.test_reaction_benchmark tests.test_motor_memory tests.test_motor_metrics tests.test_motor_benchmark tests.test_observations tests.test_observation_models tests.test_motion tests.test_motion_benchmark tests.test_transition_model tests.test_transition_pilot tests.test_transition_cli tests.test_embodiment_calibration tests.test_embodiment_cli tests.test_reference_sim -v
Expand Down Expand Up @@ -154,14 +155,6 @@ jobs:
python -m pip install -r publishing/huggingface/requirements.txt
python -m unittest discover -s publishing/huggingface -p 'test_publish_robot_sim.py' -v
python -m unittest discover -s publishing/huggingface/robot-simulation-lab -p 'test_*.py' -v
- name: Retain contact history, slowdown and recovery comparison
run: |
dvidia-piper-contact benchmark --assets runs/robot-profile-assets-ci --output runs/piper-contact-ci
dvidia-piper-contact inspect runs/piper-contact-ci
- name: Retain v0.4 recovery speed and refusal comparison
run: |
dvidia-piper-recovery benchmark --assets runs/robot-profile-assets-ci --output runs/piper-recovery-ci
dvidia-piper-recovery inspect runs/piper-recovery-ci
- name: Preserve additional profile evidence
if: always()
uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
Expand All @@ -170,12 +163,53 @@ jobs:
path: |
runs/piper-task-ci
runs/piper-grip-ci
runs/piper-contact-ci
runs/piper-recovery-ci
runs/piper-smoke-ci
runs/so101-smoke-ci
if-no-files-found: ignore

piper-comparisons:
name: PiPER ${{ matrix.cli }} / Linux / Python 3.12
runs-on: ubuntu-latest
timeout-minutes: 25
strategy:
fail-fast: false
matrix:
include:
- cli: contact
inspection: ""
- cli: recovery
inspection: ""
- cli: reliability
inspection: "--assets runs/piper-assets-ci"
steps:
- uses: actions/checkout@fbc6f3992d24b796d5a048ff273f7fcc4a7b6c09 # v5.1.0
with:
persist-credentials: false
- uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6.3.0
with:
python-version: "3.12"
- name: Provision unchanged native PiPER assets
run: |
python -m pip install ".[arm]"
dvidia-robot-sim fetch piper --output runs/piper-assets-ci
dvidia-robot-sim audit piper runs/piper-assets-ci
- name: Test new reliability decisions and receipt integrity
if: matrix.cli == 'reliability'
env:
DVIDIA_ROBOT_PROFILE_ASSETS: runs/piper-assets-ci
run: python -m unittest tests.test_piper_reacquisition tests.test_piper_grasp_strategy tests.test_piper_reliability_benchmark -v
- name: Run every planned placement and matched negative control
run: |
dvidia-piper-${{ matrix.cli }} benchmark --assets runs/piper-assets-ci --output runs/piper-comparison-ci
dvidia-piper-${{ matrix.cli }} inspect runs/piper-comparison-ci ${{ matrix.inspection }}
- name: Preserve complete comparison including failed attempts
if: always()
uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
with:
name: piper-${{ matrix.cli }}-comparison
path: runs/piper-comparison-ci
if-no-files-found: ignore

motion:
name: Causal image motion / Linux / Python 3.12
runs-on: ubuntu-latest
Expand Down
15 changes: 15 additions & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,12 @@ unloaded mechanics; learned policies and physical qualification remain open. The
[Robot Simulation Lab](https://huggingface.co/spaces/Dvidia/robot-simulation-lab)
publishes recorded evidence and local run instructions.

The [asset-library roadmap](docs/robot-simulation-library.md#reusable-asset-library-planned-product-layer)
extends this toward reusable objects, environments, grippers and sensor profiles,
starting with a shoe-organization collection. Packages will distinguish measured,
estimated and untested physical properties and link to compatible skill benchmarks.
The object catalog and installation interface are planned work.

The [PiPER grip-recovery candidate](docs/piper-grip-recovery.md) adds a short
test lift before carrying, one supported recovery attempt, and a paired
benchmark against the original controller. It uses current simulator state;
Expand All @@ -45,6 +51,15 @@ grasps are refused 2.88–2.90 seconds sooner. No completion is gained or lost.
Eight failures per mode remain visible, including two held blocks that never
satisfy the unchanged resume gate.

The [reliability v0.5 study](docs/piper-reliability-v05.md) adds a bounded second
lift and a fixed alternate-grasp planner. Three frozen fresh cohorts complete
**15/36 placements with v0.4, 18/36 with second-lift recovery, and 23/36 with the
revised grasp**, with no lost baseline successes and all 144 negative controls
failing. All 252 attempts are retained in the [complete evidence](benchmarks/piper-reliability-v05-20261012/README.md).
The revised grasp takes about 0.77 simulated seconds longer on shared successes;
thirteen failures remain. Loaded-host computation delays are reported, with no
physical real-time claim. This remains authored simulation, not a learned skill.

Train a small visual model from Skillspace footage on your own computer or server. The pipeline checks actual videos and source groups, fixes training/development/test splits, fits weights on a CPU, and exports a result with held-out measurements and integrity hashes.

The base install needs **Python 3.12+, NumPy and FFmpeg**. It needs no GPU, CUDA, account or API key. Training is offline by default; downloading software or public footage is a separate, explicit step.
Expand Down
12 changes: 12 additions & 0 deletions benchmarks/piper-reliability-development-20261011/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,12 @@
# PiPER reliability v0.5 — development archive

These experiments use already published layouts (813407 and 941827). They are development data, not fresh confirmation. Every native attempt and discarded candidate is retained with its source snapshot. The independently reserved fresh seeds are in `confirmation-reservation.json`; they were reserved before execution and must run unchanged after the selected runtime/protocol is committed.

- `lift/`: 40 attempts, 19 successes, including two rejected probe settings. The selected second lift improves known 941827 from 4/12 to 6/12 and reproduces all historical 813407 trajectories unchanged.
- `grasp/`: 76 attempts, 24 successes. Fourteen discarded aggressive-planner attempts have prohibited contacts and remain failures. The final 24 selected-candidate attempts improve known counts from 4 to 7 and 6 to 8, with zero lost successes, prohibited contacts or warnings. Shared successes are slower on average; the candidate is fixed before each episode and does not learn or select by material labels.
- `original-regressions/`: 24 exact old-v0.4 reproductions. Commands, full qpos/qvel, task metrics and outcomes all match their original receipts.
- `paired-known-941827/`: a separate 60-attempt two-mode integration, including 36 matched negative controls. Its original sealed source snapshot is in `paired-known-941827-sources/`. The final three-mode release is a different frozen protocol; inspect this intermediate bundle with its own archived source, not by substituting the later protocol.

There are **200 full native attempts** across these distinct development and regression runs. Counts include negative controls and failures; selected development success rates are not estimates of unseen performance. Unit-test and preflight fixture executions are separate. `inventory.json` seals this retained archive; original per-run receipts and seals remain untouched.

Native dynamics, model, gravity, force/gain limits, scorer and 20-second task budget remain fixed in the selected comparison. All code is authored and uses privileged simulator state. No claim of learned skills, shoe manipulation, sensor robustness or physical robot qualification follows.
Original file line number Diff line number Diff line change
@@ -0,0 +1,10 @@
{
"kind": "dvidia.piper-reliability-reservation/v0.5",
"seeds": [
832847,
227037,
862186
],
"status": "reserved, not yet executed",
"rule": "Freeze selected controller and protocol source before execution; run every cohort unchanged, no tuning between outcomes; all attempts including failures retained. Known previous cohorts remain development evidence."
}
Loading
Loading