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
89 changes: 89 additions & 0 deletions .github/workflows/continuous-grip.yml
Original file line number Diff line number Diff line change
@@ -0,0 +1,89 @@
name: Continuous shoe grip
on:
push:
pull_request:
workflow_dispatch:
permissions:
contents: read
concurrency:
group: continuous-grip-${{ github.workflow }}-${{ github.ref }}
cancel-in-progress: true
jobs:
native-development:
name: CPU development cases / Linux / Python 3.12
runs-on: ubuntu-latest
timeout-minutes: 20
env:
OPENBLAS_NUM_THREADS: "1"
OMP_NUM_THREADS: "1"
PYTHONUNBUFFERED: "1"
steps:
- uses: actions/checkout@fbc6f3992d24b796d5a048ff273f7fcc4a7b6c09
with:
persist-credentials: false
- uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1
with:
python-version: "3.12"
- uses: actions/setup-node@49933ea5288caeca8642d1e84afbd3f7d6820020
with:
node-version: "22"
- name: Reproduce the pinned replay viewer
run: |
npm ci --ignore-scripts --prefix viewer
npm run --prefix viewer build:grip
git diff --exit-code -- src/dvidia_training/grip_lab.bundle.js
- name: Provision pinned native PiPER mechanics
run: |
python -m pip install '.[arm]'
python -m pip check
dvidia-grip-benchmark --help
dvidia-robot-sim fetch piper --output runs/prerequisite-assets-ci
dvidia-robot-sim audit piper runs/prerequisite-assets-ci
- name: Check native development cases and evidence contracts
env:
DVIDIA_ROBOT_PROFILE_ASSETS: runs/prerequisite-assets-ci
run: python -m unittest tests.test_continuous_grip tests.test_grip_protocol tests.test_grip_monitor_evidence tests.test_grip_benchmark -v
- name: Retain every development attempt and its replay
run: >-
dvidia-grip-benchmark --assets runs/prerequisite-assets-ci
--output runs/grip-development-ci --split development
--source-commit "$GITHUB_SHA"
- name: Verify exported receipts and checksums
run: |
python - <<'PY'
from hashlib import sha256
import json
from pathlib import Path
root = Path('runs/grip-development-ci')
protocol = json.loads((root / 'protocol.json').read_text())
report = json.loads((root / 'benchmark.json').read_text())
assert protocol['split'] == report['split'] == 'development'
expected = {row['id'] for row in protocol['attempts']}
assert expected == {row['id'] for row in report['attempts']}
assert len(report['attempts']) == len(expected)
assert all(row['same_compiled_model'] for row in report['paired_comparisons'])
assert all(row['valid'] for row in report['attempts']), 'A native or export error is not an expected task failure'
assert (root / 'benchmark.json').stat().st_size <= 8 * 1024 * 1024, 'Viewer index exceeds its fetch limit'
for row in report['attempts']:
receipt = json.loads((root / 'trials' / (row['id'] + '.json')).read_text())
assert receipt['case']['split'] == 'development'
assert receipt['success'] == row['success'] and receipt['valid'] == row['valid']
assert receipt['native_trace'] and receipt['native_trace_columns']
times = [sample[0] for sample in receipt['native_trace']]
assert all(abs((b - a) - .002) < 1e-8 for a, b in zip(times, times[1:])), 'Dense trace lost a native step'
assert all((root / value).is_file() for value in row['replay'].values())
integrity = json.loads((root / 'integrity.json').read_text())
for name, expected_file in integrity['files'].items():
content = (root / name).read_bytes()
assert len(content) == expected_file['bytes']
assert sha256(content).hexdigest() == expected_file['sha256'], name
print(json.dumps({'attempts_retained': len(expected), 'groups': report['groups'],
'scope': 'Development regression only; no heldout run or hardware qualification.'}))
PY
- name: Retain evidence even when a check fails
if: always()
uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02
with:
name: continuous-grip-development
path: runs/grip-development-ci
if-no-files-found: ignore
10 changes: 10 additions & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,16 @@

**Research alpha.** The recorded demo covers the original one-scene simulation pilot; the frozen benchmark is a separate experiment. Training runs locally. No human-video action bridge or hardware skill is established.

**[Continuous shoe grip v0.2](https://research.dvidia.org/papers/continuous-shoe-grip-v02/).**
The mounted-rake experiment now checks grip throughout motion, tests a small
lift and lowers onto support before a bounded regrasp. The fresh comparison
completes **2/10 lifts under both controllers**, on the same cases. Twelve
lowering responses across six failed grasp cases add time without recovering a
completed lift. All twenty attempts, including failures, are available in the
[recorded comparison](https://dvidia-robot-simulation-lab.static.hf.space/prerequisites/v02/index.html).
[Run locally](docs/continuous-shoe-grip.md) ·
[Frozen results and independent audit](benchmarks/continuous-shoe-grip-20261013/README.md).

**[Explore the two skill-prerequisite experiments](https://research.dvidia.org/papers/skill-prerequisites-v01/).**
The hinged shoe box and out-of-reach shoe now have runnable native MuJoCo scenes,
recorded 3D replays and all sixteen held-out attempts. A lid guard avoids two
Expand Down
45 changes: 45 additions & 0 deletions benchmarks/continuous-shoe-grip-20261013/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,45 @@
# Continuous shoe grip v0.2 — frozen comparison

Source: [`af8933994b4626e28a0554abc782ae9d22cff06c`](https://github.com/Dvidia-Inference/dvidia-training/tree/af8933994b4626e28a0554abc782ae9d22cff06c).

The fresh comparison retains **20/20 valid native attempts**. Both authored controllers complete **2/10**, on the same two cases; there are no gains or losses. Both enter the retrieval region in 8/10 and correctly defer two short-tool cases. The new controller issues 12 bounded lowering responses across six unsuccessful grasp cases. Those cycles add 34.02 simulated seconds and produce no additional completed tasks.

[Research paper](https://research.dvidia.org/papers/continuous-shoe-grip-v02/) · [Recorded comparison and full receipts](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-10-13-continuous-shoe-grip-v02/lab/) · [Hugging Face](https://huggingface.co/spaces/Dvidia/robot-simulation-lab) · [Run locally](../../docs/continuous-shoe-grip.md)

## Scope and evidence

`baseline` is the original **mounted-rake** v0.1 controller. It is not the earlier no-tool capability control. All ten paired compiled-model hashes agree. The new controller uses continuous acquisition/loss checks, a 40 mm observed TCP checkpoint during continued lifting, and a bounded supported retry. The checkpoint is not independent shoe-lift success.

- [Fresh benchmark index](benchmark.json), [protocol](protocol.json), [full export inventory](integrity.json).
- [Historical regression index](regression-benchmark.json), [protocol](regression-protocol.json), [source export inventory](regression-integrity.json): 18/18 valid attempts, 1/9 completed under either controller on the same known development case. These are previously seen cases.
- [Development ledger](development-ledger.json): all nine tuning rounds are summarized. First-five source hashes were not recorded; no retroactive provenance is claimed. Its first iteration's peak height included initialization and is not an earned lift.
- [Timing and outcome derivation](analysis.json).
- [Independent native audit](native-audit.md), [machine-readable findings](native-audit.json), [portable reproduction script](native_audit.py) and [local release acceptance](acceptance.json).

This directory contains compact results, not the full replay exports. `integrity.json` inventories the complete fresh 87-file export on the research site and Hugging Face. Regression replay paths refer to the original local export; regression trajectories are not included in this compact directory. The development ledger refers to local historical iteration artifacts by name and hash, rather than bundling those full trajectories.

The independent audit verifies nine bilateral-contact-loss alarms at **20–52 ms, median 30 ms**, after the latest uninterrupted native contact-loss segment. Three separate acquisition-failure alarms have no acquired-grip loss origin and are excluded. Three failed 20 ms samples can straddle brief native contact reacquisition. Recorded request-to-packet timing is not physical stopping or response time. All twelve lowering responses establish world support before reopening; none completes afterward.

## Remaining failures

The independent scorer requires the shoe's sole-bottom frame above 35 mm, both finger contacts above 0.01 N, and world/rake/non-finger robot support below 0.01 N for every tick of the final 500 ms. A tilted high pose or earlier good hold cannot pass by itself.

Both policies have an uncovered non-finger-support interval in completed case `fresh-05`: 82 failed ticks from 19.076–19.238 s, with 100 ms requalification complete at 19.338 s. This is support contamination, not evidence of physical slipping. Neither controller detects it; the final window passes after it clears. All 12 new controller detections are unmatched to the independent loss witness, which requires prior 100 ms qualifying retention. They are not automatically false alarms or measured missed-slip rates.

The [geometry probe summaries](geometry-probes/probe.json), including [the lower-envelope diagnostic](geometry-probes/lower-envelope.json), retain 32 unsuccessful probes across seven files. The diagnostic lowered the software TCP floor to 35 mm; the release retains its original 55 mm floor. One diagnostic reached a 181.6 mm shoe frame height while carrying 1.816 N of rake support, so it correctly failed. Finite probes do not prove grasp infeasibility; they motivate examining gripper clearance and the mounted tool's interference before adding more retries.

## Compute

On Apple M5 / 10 logical CPUs / 24 GiB RAM, macOS arm64, Python 3.12.9, MuJoCo 3.15.0:

| Measurement | v0.1 mounted-rake baseline | v0.2 recovery |
| --- | ---: | ---: |
| Simulated task duration, summed | 131.60 s | 165.62 s |
| Native execution wall time | 4.9601 s | 5.7716 s |
| Separate audit and export wall time | 8.9476 s | 8.7451 s |
| Recorded target-solver calls | 6,580 | 8,281 |
| Target-solver p95 / maximum | 0.0305 / 0.5635 ms | 0.0316 / 0.3753 ms |

Native timing includes setup, control, physics and recording. Target-solver samples exclude most monitoring, contact observation, physics, network and hardware latency. None of the recorded solver calls exceeds 20 ms; that is not a whole-loop or hard-real-time guarantee. No GPU is required for this native benchmark. Browser replay needs WebGL2; minimum RAM is unmeasured.

These are authored policies with privileged observations, rigid synthetic shoes and assumed contact. No footage-trained model, force calibration or physical-robot qualification is claimed. Code and synthetic assets are MIT; upstream notices remain with the complete export.
39 changes: 39 additions & 0 deletions benchmarks/continuous-shoe-grip-20261013/acceptance.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,39 @@
{
"schema": "dvidia.continuous-grip-publication-acceptance/v1",
"source_commit": "af8933994b4626e28a0554abc782ae9d22cff06c",
"tests_passed": true,
"browser_qa_passed": true,
"benchmark_reviewed": true,
"benchmark_sha256": "cf909bdf80bd29af7627b312756e0aaba9c77283e353ff300747c28d86bc4431",
"integrity_sha256": "827cd21905080fb6e60a93484d76df095ff74342c09b399136d6906709bd5998",
"native_audit_sha256": "51114b4adc79db7ff8c095d42fd32ccc6c5c71adea402ea2e073e363df1e5cfc",
"test_evidence": {
"native_and_reporting_tests": 45,
"publisher_tests": 19,
"log": "runs/grip-v02-dev/final-tests.log"
},
"browser_checks": {
"frozen_export_native_3d_render": true,
"completed_fresh08_both_policies": true,
"failed_fresh01_recovery_event_seek": true,
"short_tool_fresh09_both_policies_defer": true,
"completed_playback_restarts_and_pauses": true,
"keyboard_timeline_end": true,
"collision_toggle_and_reset_on_case_change": true,
"event_phase_filter": true,
"responsive_preflight_320px_no_horizontal_overflow": true,
"responsive_preflight_390px_no_horizontal_overflow": true,
"responsive_scope": "UI preview of same layout before source freeze; frozen desktop release then rechecked.",
"screenshot": "frozen-viewer.jpg"
},
"result_review": {
"valid_attempts": 20,
"baseline_completed": 2,
"continuous_grip_completed": 2,
"paired_gains": 0,
"paired_losses": 0,
"completed_after_recovery": 0,
"limitations_acknowledged": true
},
"publication_scope": "Gate for exact frozen lab export. Public deployment behavior is checked separately after release."
}
73 changes: 73 additions & 0 deletions benchmarks/continuous-shoe-grip-20261013/analysis.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,73 @@
{
"schema": "dvidia.continuous-shoe-grip-analysis/v2",
"source_commit": "af8933994b4626e28a0554abc782ae9d22cff06c",
"benchmark_sha256": "cf909bdf80bd29af7627b312756e0aaba9c77283e353ff300747c28d86bc4431",
"regression_benchmark_sha256": "fd3a6ca2fb8d58c8faf32a553862ad70029f39d5adf33b76096b32ccb56c619b",
"host_supplement": {
"cpu": "Apple M5",
"logical_cpus": 10,
"memory_bytes": 25769803776,
"source": "Root host inventory; shared-host measurements, not minimum requirements."
},
"policies": {
"baseline": {
"attempts": 10,
"completed": 2,
"retrieval_evidenced": 8,
"appropriate_deferrals": 2,
"simulated_seconds": 131.60000000000312,
"native_wall_seconds": 4.960062542697415,
"simulated_to_native_wall_ratio": 26.53192351248771,
"audit_export_wall_seconds": 8.947619959013537,
"target_solver_calls": 6580,
"target_solver_p50_ms": 0.02208305522799492,
"target_solver_p95_ms": 0.030502211302518813,
"target_solver_max_ms": 0.5634580738842487,
"target_solver_calls_over20ms": 0,
"recovery_requests": 0,
"issued_recovery_packets": 0,
"sustained_condition_losses": 1,
"condition_losses_without_detection": 1,
"unmatched_detection_count": 0
},
"recovery": {
"attempts": 10,
"completed": 2,
"retrieval_evidenced": 8,
"appropriate_deferrals": 2,
"simulated_seconds": 165.61999999999634,
"native_wall_seconds": 5.77158858301118,
"simulated_to_native_wall_ratio": 28.695739070436012,
"audit_export_wall_seconds": 8.745075207436457,
"target_solver_calls": 8281,
"target_solver_p50_ms": 0.022500054910779,
"target_solver_p95_ms": 0.031624920666217804,
"target_solver_max_ms": 0.3752908669412136,
"target_solver_calls_over20ms": 0,
"recovery_requests": 12,
"issued_recovery_packets": 12,
"sustained_condition_losses": 1,
"condition_losses_without_detection": 1,
"unmatched_detection_count": 12
}
},
"timing_scope": "Native outer timer includes setup, control, physics and recording; separate audit/export excludes startup and final report write. Target-solver samples exclude sensing, most monitoring, physics and hardware latency. No hard-real-time claim.",
"paired_gains": 0,
"paired_losses": 0,
"same_compiled_models": 10,
"additional_simulated_seconds": 34.01999999999322,
"fresh05_condition_gap": {
"policies": [
"baseline",
"recovery"
],
"onset_s": 19.076,
"last_failed_tick_s": 19.238,
"failed_ticks": 82,
"failed_tick_duration_s": 0.164,
"requalified_s": 19.338,
"cause": "non-finger robot support >=0.01 N; other final-retention predicates remain valid throughout this interval",
"detected": false,
"scope": "Support contamination, not demonstrated physical slipping; final window succeeds after support clears."
}
}
Loading
Loading