Skip to content
Merged
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
  •  
  •  
  •  
8 changes: 8 additions & 0 deletions .gitattributes
Original file line number Diff line number Diff line change
Expand Up @@ -6,3 +6,11 @@
**/2026-10-11-native-grasp/replay/receipt/** linguist-generated=true -diff
**/2026-10-11-native-grasp/replay/motion.json linguist-generated=true -diff
**/2026-10-11-native-grasp/replay/scene.json linguist-generated=true -diff

# Preserve exact generated MuJoCo/Three.js runtime bytes and upstream shaders.
**/workstation/*.bundle.js -diff linguist-generated=true
**/workstation/mujoco.wasm binary
**/2026-10-13-live-piper-workstation-v01/workstation/assets/** -diff linguist-vendored=true
**/2026-10-13-live-piper-workstation-v01/evidence/**/attempts.json -diff linguist-generated=true
**/2026-10-13-live-piper-workstation-v01/evidence/browser/browser-raw-samples.json -diff linguist-generated=true
**/2026-10-13-live-piper-workstation-v01/evidence/browser/session-receipt.json -diff linguist-generated=true
14 changes: 14 additions & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,20 @@ All four covers are **AI-generated editorial illustrations**, not footage, measu

## Latest studies


[Live PiPER shoe workstation v0.1](reports/Live%20PiPER%20shoe%20workstation%20v01.md)
lets readers drag a tool target, adjust the gripper and nudge a free rigid shoe
while MuJoCo runs locally in the browser. Native physics, bounded authored
control and live diagnostics now accompany the recorded asset previews.
The final native run retains **20 valid attempts**, with **14/16** target trials
passing and both high-target repetitions failing. Jaw, nudge, gravity and a
600-second native hold pass. Development and native/browser comparisons remain
separate. All fifteen development shoe-lift attempts failed;
retained lifting, learned shoe skills and hardware transfer remain unqualified.
[Open the live workstation](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-10-13-live-piper-workstation-v01/workstation/index.html)
or [run the Hugging Face copy](https://dvidia-robot-simulation-lab.static.hf.space/workstation/v01/index.html).


[PiPER placement reliability v0.5](reports/PiPER%20reliability%20v05.md)
completes **23/36 placements**, versus **15/36** for frozen v0.4 and **18/36**
with second lift alone. Three fresh cohorts retain **252 valid attempts** and
Expand Down
146 changes: 146 additions & 0 deletions docs/downloads/reports/Live PiPER shoe workstation v01.md

Large diffs are not rendered by default.

14 changes: 14 additions & 0 deletions docs/downloads/topics/physical-grounding/README.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,19 @@
# Physical grounding


## Live PiPER shoe workstation — 13 October 2026

[Live PiPER shoe workstation v0.1](../../reports/Live%20PiPER%20shoe%20workstation%20v01.md)
adds browser-local MuJoCo physics, a fixed-orientation target controller, native
jaw commands and a free rigid trainer on a physical worktop. Users can move a
target, inspect tracking and contact, nudge the shoe and download session
diagnostics. The [run archive](runs/2026-10-13-live-piper-workstation-v01/README.md)
keeps frozen benchmarks, exploratory failures and native/browser comparisons
separate. All fifteen development retained-lift attempts failed. Shoe properties
remain untested assumptions; no learned shoe policy or hardware qualification is
established. [Open the live workstation](runs/2026-10-13-live-piper-workstation-v01/workstation/index.html).


## Robotics research pipeline — 9 October 2026

[Robotics research pipeline](../../reports/Robotics%20research%20pipeline.md) connects footage review, synchronized measured state, separate visual and movement training, frozen simulation evaluation and versioned skill installation. Its diagram identifies current simulation components, proposed automatic labeling and learned task decisions, and the outstanding physical qualification gate. The proposed supervisor consumes recent joint, gripper and contact evidence; it does not equate a close command with a successful grasp. The new annotation-labor pilot is a companion to the existing graduation proposal. Both new experiments are **not run**. Existing recorded benchmarks retain their original protocols and claims.
Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,19 @@
# Live PiPER shoe workstation v0.1

[Open the live workstation](workstation/index.html) · [Research report](../../../../reports/Live%20PiPER%20shoe%20workstation%20v01.md) · [Hugging Face](https://huggingface.co/spaces/Dvidia/robot-simulation-lab) · [GitHub implementation](https://github.com/Dvidia-Inference/dvidia-training/pull/17)

Drag the tool target, actuate the native gripper and nudge a free rigid trainer. MuJoCo runs locally in the browser with 500Hz physics and 50Hz authored control. The 99-file export works offline after provisioning and local HTTP serving.

**14/16 target attempts passed; 20/20 attempts were numerically valid.** Both high-target repeats failed. Jaw travel, shoe impulse, removed-table gravity and 600 simulated seconds of native holding passed. The 15 development grasp attempts all failed retained lift. This release does not contain a trained shoe-organizing policy or hardware qualification.

The [final ten-minute browser run](evidence/browser/browser-benchmark.json) reports **51.39 render calls/s** and a **0.99998 simulation/wall-time ratio** on the tested Apple M5 / Chromium 151 host, with no reported lag or warnings. Simulated TCP movement response is **61.9 ms p95** across 61 commands. Render calls do not measure GPU presentation or robot actuator latency.

The [native/WASM comparison](evidence/parity/wasm-parity.json) passes numerical state and loaded-contact-count checks. **Strict raw-contact-count parity fails in 13/1,500 states**, and the [nonzero exit status](evidence/parity/run-status.json) is retained.

[Complete evidence, failures and screenshots](evidence/README.md) · [Native report](evidence/native/report.json) · [Sealed simulator inventory](workstation/workstation-integrity.json) · [Source binding](evidence/source-binding.json)

Frozen runtime: [79ffdda1c0759bce1201ce3391c40403c385c39b](https://github.com/Dvidia-Inference/dvidia-training/tree/79ffdda1c0759bce1201ce3391c40403c385c39b). Publication tooling and reproducible browser-check instructions are available in the linked PR. Source binding preserves the original receipt's null commit field and proves every recorded implementation hash matches the frozen revision.

Shoe material properties are synthetic assumptions. The visible laces and fabric are cosmetic; collision geometry remains rigid. GPU rendering, native CPU timing, state publication and actual robot latency are separate quantities.

DVIDIA code and original assets: MIT. PiPER simulation assets and Three.js: MIT. MuJoCo: Apache-2.0 with dependency notices preserved in the export. Original research writing: CC BY 4.0.
Original file line number Diff line number Diff line change
@@ -0,0 +1,21 @@
# Live PiPER shoe workstation v0.1 — complete evidence

Frozen runtime: [79ffdda1c0759bce1201ce3391c40403c385c39b](https://github.com/Dvidia-Inference/dvidia-training/tree/79ffdda1c0759bce1201ce3391c40403c385c39b).

Native confirmation: **14/16 tracking attempts passed; all 20 attempts numerically valid**. Both high-target repeats failed, with 31.44mm final error. Jaw travel, physical shoe nudge, disabled-table gravity and 600 simulated seconds of native holding passed. This is authored control of a rigid synthetic shoe; retained shoe lifting is unqualified and physical robot transfer is unmeasured.

- [Native report](native/report.json), [frozen protocol](native/protocol.json), [all attempts and trajectories](native/attempts.json), and [native inventory](native/integrity.json).
- [Source binding](source-binding.json): native run began immediately before the Git commit. Its original null source-commit field remains unchanged; every protocol implementation hash matches the frozen revision.
- [UI acceptance](ui-acceptance/acceptance.json), [session receipt](ui-acceptance/session.json), [desktop screenshot](ui-acceptance/workstation-1440.png), [mobile screenshot](ui-acceptance/workstation-320.png).
- [Installed-wheel validation](wheel-validation.json): all 99 exported files match the frozen source export byte for byte.
- [All 15 failed development grasp attempts](development/grasp-search.json), [orientation exploration](development/orientation-search.json), and [four controller-cap trajectories](development/controller-caps-search.json).
- Earlier development [baseline report](development/native-benchmark/report.json) and [candidate report](development/native-candidate/report.json), including the invalid old push request and both high-target failures. Complete attempt records and integrity files accompany each.
- Preliminary [native comparison](development/native-parity-30s.json) and [strict WASM comparison failure](development/wasm-parity-30s.json), plus the contact-force audit. Raw near-zero contact-count differences remain failures of strict parity.
- Preliminary software-rendered [browser result](development/browser-dev-v05-20s-r2/browser-benchmark.json) and original [form-draft failure](development/browser-dev-20s/browser-failure.json). Development latency quantiles used an earlier estimator; final browser evidence uses nearest-rank quantiles. Software rasterization does not measure the laptop GPU.

- [Final 600-second browser benchmark](browser/browser-benchmark.json), [raw samples and durations](browser/browser-raw-samples.json), and [bounded session receipt](browser/session-receipt.json). Apple M5 Metal / Chromium 151: 51.39 render invocations/s, 0.99998× real time, no warnings or accumulated lag. Across 61 target commands, target echo p95=20.6ms and first observed>=1mm TCP movement p95=61.9ms; all responded. Rendering calls are not measured GPU presentation completions, and response is simulator/UI latency, not hardware latency.
- [Final native fixture](parity/native-parity.json) and [WASM comparison](parity/wasm-parity.json) include an off-control-boundary gripper command. Numerical and loaded-contact comparisons pass; strict raw counts differed in 13/1,500 states (extra forces<=4.44e-10N), so strict parity failed and the nonzero exit code is retained.

The development browser failure's exception stack has its local checkout prefix redacted in the public copy, with the original file hash recorded. All numerical results remain unchanged. Native timing includes control, ten physics steps, forward kinematics and diagnostics; browser rendering and hardware actuation are excluded. Tests ran on a shared development laptop, not an isolated real-time host.

DVIDIA software, original shoe geometry and screenshots: MIT. MuJoCo: Apache-2.0 with dependency notices; PiPER simulation assets and Three.js: MIT. Research prose: CC BY 4.0. Keep upstream notices with the full workstation export.
Original file line number Diff line number Diff line change
@@ -0,0 +1,183 @@
{
"schema": "dvidia.workstation-browser-benchmark/v1",
"passed": true,
"passed_scope": "Functional UI acceptance and numerical validity; performance targets are reported separately.",
"functional_acceptance_passed": true,
"quantile_method": "Nearest-rank empirical quantiles; sample counts are reported for each distribution.",
"run_started_at": "2026-10-13T07:22:34.716Z",
"url": "http://127.0.0.1:8295/browser-final/index.html",
"manifest_sha256": "6764027bf57e61891c6f60a28abb1e354975e6174c9092f7d7c00fce39a58d5f",
"native_scene_sha256": "3400d5e68731ce032358a538a67eb87e00a0613620553988d2f53c851641328b",
"engine": "MuJoCo 3.15.0 WASM",
"browser": "151.0.7922.34",
"browser_channel": "chromium",
"headless": true,
"requested_renderer": "device default",
"renderer": {
"vendor": "Google Inc. (Apple)",
"renderer": "ANGLE (Apple, ANGLE Metal Renderer: Apple M5, Unspecified Version)",
"version": "WebGL 2.0 (OpenGL ES 3.0 Chromium)",
"shading_language": "WebGL GLSL ES 3.00 (OpenGL ES GLSL ES 3.0 Chromium)",
"user_agent": "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) HeadlessChrome/151.0.0.0 Safari/537.36",
"hardware_concurrency": 10,
"device_memory_gib": 16
},
"renderer_classification": "renderer reported above; hardware backing not independently established",
"host": {
"os": "darwin 25.5.0 arm64",
"cpu": "Apple M5",
"logical_cpus": 10,
"ram_bytes": 25769803776
},
"viewport": {
"width": 1440,
"height": 1080,
"device_scale_factor": 1
},
"load_to_ready_ms": 9215.773917,
"controls": {
"target_motion_m": 0.004066944147514631,
"target_apply": true,
"gripper_closed_m": 0.009231686038699782,
"gripper_open_m": 0.025209206867549547,
"pause_preserved_sim_time": true,
"reset_restored_qpos": true,
"nudge_displacement_m": 0.004635949741912616,
"nudge_physical_response": true,
"receipt_export": true,
"receipt_retention": {
"sample_wall_interval_ms": 100,
"sample_limit": 3000,
"command_limit": 2000,
"received_states": 31246,
"samples_dropped": 2406,
"commands_dropped": 0,
"note": "Bounded recent samples and commands; wall timestamps continue across reset. Simulation time restarts on reset. UI samples may omit intervening physics steps."
}
},
"numerical_validity": true,
"performance_targets": {
"render_30_to_60_fps_met": true,
"real_time_physics_met": true,
"note": "Performance depends on the recorded renderer, device and browser; failed targets do not change functional acceptance."
},
"stability": {
"requested_wall_seconds": 600,
"measured_wall_seconds": 600.0062999999523,
"simulated_seconds": 599.9939999996213,
"real_time_factor": 0.9999795002146962,
"physics_steps": 299997,
"lag_added_ms": 0,
"foreground_throughout": true,
"foreground_definition": "document.visibilityState remained visible in the isolated browser; headless runs have no OS-visible window.",
"warning_count": 0,
"native_state_samples": 5352,
"cyclic_target_interval_seconds": 10,
"cyclic_target_commands": 60
},
"rendering": {
"render_invocations": 30834,
"render_invocations_per_second": 51.38946041066978,
"render_interval_ms": {
"count": 30833,
"p50": 16.799999952316284,
"p95": 34.200000047683716,
"p99": 35.799999952316284,
"min": 5.799999952316284,
"max": 43.299999952316284
},
"raf_callbacks_per_second": 59.999370006619706,
"raf_interval_ms": {
"count": 35999,
"p50": 16.699999999982538,
"p95": 18.5,
"p99": 18.70000000001164,
"min": 14.5,
"max": 18.800000000017462
},
"worker_states_per_second": 51.569458520689636,
"worker_state_interval_ms": {
"count": 30941,
"p50": 19.40000009536743,
"p95": 21.700000047683716,
"p99": 22.899999856948853,
"min": 0.7000000476837158,
"max": 35.5
},
"target_30_to_60_render_fps_met": true,
"note": "RAF cadence, Three.js render invocations and worker-state delivery are separate measurements. Render invocations do not measure GPU presentation completion."
},
"response_latency": {
"measurement": "Main-thread command post to first received target echo, separately from first >=1 mm TCP movement projected toward the target.",
"stable_baseline": "At least 5 recent states within 350 ms, spread below 0.5 mm; commanded distance exceeds 5 mm.",
"response_timeout_ms": 5000,
"superseded_commands": "Unfinished responses remain missing when a newer target is issued.",
"all_target_commands": 61,
"stable_baseline_commands": 61,
"target_echo_ms": {
"count": 61,
"p50": 13.900000095367432,
"p95": 20.59999990463257,
"p99": 46.5,
"min": 5.199999809265137,
"max": 46.5
},
"physical_response_ms": {
"count": 61,
"p50": 54.60000014305115,
"p95": 61.89999985694885,
"p99": 70.70000004768372,
"min": 36.5,
"max": 70.70000004768372
},
"missing_physical_response": 0
},
"observed_worker_timing": {
"step_ms_p95": {
"count": 5352,
"p50": 0.2999999523162842,
"p95": 0.39999985694885254,
"p99": 0.40000009536743164,
"min": 0.10000014305114746,
"max": 0.5
},
"control_ms_p95": {
"count": 5352,
"p50": 0.20000004768371582,
"p95": 0.2999999523162842,
"p99": 0.2999999523162842,
"min": 0.10000014305114746,
"max": 0.2999999523162842
},
"note": "Distribution of worker rolling p95 snapshots, not an all-step timing distribution."
},
"memory": {
"browser_heap": {
"used_js_heap_size": 48572857,
"total_js_heap_size": 113567557,
"js_heap_size_limit": 4395630592
},
"receipt_samples": 3000,
"receipt_sample_limit": 3000,
"receipt_samples_dropped": 2406,
"harness_states_retained": 5393,
"harness_state_limit": 15000,
"harness_states_dropped": 0,
"harness_frame_samples_dropped": 0,
"worker_timing_samples_retained": 2048
},
"request_origins": [
"http://127.0.0.1:8295"
],
"errors": [],
"console_errors": [],
"failed_requests": [],
"response_errors": [],
"unexpected_requests": [],
"warnings": [],
"limitations": [
"Rigid synthetic shoe with uncalibrated materials.",
"No shoe-grasp or real-robot skill success claim.",
"Browser instrumentation adds overhead; software rendering results do not establish device-GPU performance."
]
}
Loading
Loading