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
Jump to file
Failed to load files.
Loading
Diff view
Diff view
35 changes: 34 additions & 1 deletion .github/workflows/ci.yml
Original file line number Diff line number Diff line change
Expand Up @@ -67,9 +67,42 @@ jobs:
dvidia-observation-studio --help
dvidia-motion --help
dvidia-transition --help
dvidia-embodiment --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 -v
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

reference:
name: Hardware reference / Linux / Python 3.12
runs-on: ubuntu-latest
timeout-minutes: 15
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 pinned reference models
run: |
python -m pip install ".[arm]"
dvidia-embodiment fetch-assets --output runs/reference-assets-ci --hands
dvidia-embodiment verify-assets runs/reference-assets-ci --hands
- name: Test compiled mechanics, collisions and replay scoring
env:
DVIDIA_REFERENCE_ASSETS: runs/reference-assets-ci
run: python -m unittest tests.test_reference_sim tests.test_embodiment_cli tests.test_embodiment_calibration -v
- name: Retain all numerical reference attempts
run: |
dvidia-embodiment benchmark runs/reference-assets-ci --hands --output runs/reference-bench-ci --repeats 1 --seconds 2
dvidia-embodiment inspect runs/reference-bench-ci
- name: Preserve reference evidence
if: always()
uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
with:
name: hardware-reference-bench
path: runs/reference-bench-ci
if-no-files-found: ignore

motion:
name: Causal image motion / Linux / Python 3.12
Expand Down
8 changes: 8 additions & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -119,6 +119,14 @@ Recording quotas are proposed study sizes, not known sufficiency thresholds.

## Try the pipeline

The [hardware reference bench](docs/hardware-reference.md) assembles a pinned
UR5e arm and coupled Robotiq gripper, audits actual joint/actuator/collision
structure, and compares CPU physics at three timesteps. Optional left/right
five-finger references have separate contracts. A paired replay scorer measures
joint, tool and jaw errors once aligned physical recordings are available.
These tools use the `feat/hardware-reference-v0.1` candidate branch; the new
reference has no learned task policy or measured physical fidelity.

Generate an explicitly synthetic example, then train and verify its output:

```sh
Expand Down
21 changes: 21 additions & 0 deletions benchmarks/hardware-reference-20261011/LICENSE
Original file line number Diff line number Diff line change
@@ -0,0 +1,21 @@
MIT License

Copyright (c) 2026 DVIDIA contributors

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
37 changes: 37 additions & 0 deletions benchmarks/hardware-reference-20261011/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,37 @@
# Hardware reference mechanics — 11 October 2026

Frozen CPU diagnostics from runtime [c95a438](https://github.com/Dvidia-Inference/dvidia-training/commit/c95a4385dae4b654b01ef75f56cbcc144fde2f5c). All **18/18 attempted runs are numerically valid**, with no native warnings. These are deterministic unloaded motion/closure experiments, not eighteen independent skill trials or physical qualification.

The compiled UR5e/Robotiq reference has fourteen scalar joint coordinates, seven actuator commands, three equality objects, one split tendon and 24 collision geoms. All six moving arm links have active collision proxies; the upstream fixed base has none. Both separately audited Shadow references have 24 joint coordinates, 20 commands and four summed-angle tendons. The bimanual assembly is not implemented. No native sensor channels, learned task policy, task object or physical calibration is supplied.

## CPU cost at a 1 ms physics step

| Authored motion | Throughput, two repetitions | 20 ms chunk p95, two repetitions (ms) |
| --- | --- | --- |
| hold_open | 44.41×, 47.29× | 0.574125, 0.562250 |
| arm_sweep | 44.32×, 40.80× | 0.711583, 0.708542 |
| jaw_cycle | 41.79×, 42.83× | 0.685666, 0.682750 |

Hardware: macOS-26.5.2-arm64-arm-64bit, Python 3.12.9, MuJoCo 3.15.0. No GPU used. Throughput is simulated duration divided by native stepping/instrumentation/boundary-forward wall time; it excludes verification/compile, settling, sampling, rendering, perception, policy, communication and physical actuation. It is not total training throughput, physical reaction speed or a measured minimum machine. CPU load/power state may change these timings.

## Numerical sensitivity and contact coverage

Each of three cases runs four seconds with 2, 1 and 0.5 ms native timesteps and two repeats. Issued commands stay fixed at 20 ms boundaries. Adjacent-step comparisons retain all 201 matching boundaries, including post-settling starts. For arm sweep, TCP translation RMSE drops from 0.074238 mm (2→1 ms) to 0.037023 mm (1→0.5 ms). For jaw cycling, pad-center separation RMSE drops from 0.034480 mm to 0.017331 mm. This compares simulation with simulation, not reality, and does not prove solver convergence across contact-rich tasks.

The static sphere-overlap diagnostic records one upper-arm contact pair with ordinary masks and zero with arm masks disabled. It tests geometry participation, not avoidance, stopping or physical contact accuracy. Pad-center separation is a geometric proxy, not inner jaw gap or calibrated force. Repeated samples/cases are correlated.

## Artifacts and reproduction

- `report.json`: settings, source hashes, every attempted run, full native trajectories, inventory, collision diagnostic and numerical comparisons.
- `report.html`: offline schematic replay and errors/timing; no remote dependencies or physical commands.
- `integrity.json`: exact hashes/sizes for the two report artifacts; internal consistency only.
- `source.json`: supplemental post-run source attribution, outside the artifact seal.
- `LICENSE`: MIT for DVIDIA's original software/report. Third-party assets are not included in this bundle; their BSD-3-Clause/BSD-2-Clause/Apache-2.0 source paths and digests are retained in the reference lock.

```sh
.venv/bin/dvidia-embodiment inspect benchmarks/hardware-reference-20261011
.venv/bin/dvidia-embodiment fetch-assets --output runs/reference-assets --hands
.venv/bin/dvidia-embodiment benchmark runs/reference-assets --hands --output runs/new-reference-bench
```

[Install and calibration protocol](../../docs/hardware-reference.md). Physical mismatch can be scored from complete aligned physical/simulated traces; the included analytic paired fixture is synthetic and is not this arm's hardware data. Images alone do not supply calibrated geometry, actuator response, grip/friction, actual sensing or a trained action policy. LeRobot can be the future collection/device layer while DVIDIA builds its own model; it is not integrated into this version.
13 changes: 13 additions & 0 deletions benchmarks/hardware-reference-20261011/integrity.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,13 @@
{
"schema": "dvidia.reference-integrity/v1",
"files": {
"report.json": {
"bytes": 10258558,
"sha256": "f619206af4580dff23cca67dafbe227788a4b5f1acf1efe64b0e8de249a1e5e1"
},
"report.html": {
"bytes": 5207740,
"sha256": "c9f5a352fc365f23ed0fb52467d090f46ef63a4825250fa1bf56f77f867c2cda"
}
}
}
28 changes: 28 additions & 0 deletions benchmarks/hardware-reference-20261011/report.html

Large diffs are not rendered by default.

Loading
Loading