# A local camera-to-action learner reaches 5/64 strict placements

**Status: the completed study retains the model as a research candidate.**
The selected learner completed **5/64 strict unseen-start placements**, versus
9/64 for the matched blind learner, 31/64 for the fixed primitive and 64/64 for
the authored teacher. The sealed report fails the registered qualification
gates. Independent full-study audit passed on 9 November 2026 with no errors;
this confirms the saved evidence, not skill qualification. No qualified
downloadable skill or physical robot readiness is claimed.

## Question and bounded experiment

Can a small local model learn a complete PiPER block-placement task from
synchronized RGB-D, measured robot state and explicit synthetic action labels?
The intended learned sequence is approach, grasp, lift, carry, release and
retreat. The study uses one simulated PiPER, its existing gripper, a known rigid
40 mm block and two fixed destinations. It does not recover motor actions from
ordinary human video.

All starts, controller assignments, training settings and acceptance gates were
registered before collection. The runtime is frozen at
`c9a7c55531533f60e24651a99d509f721e195051`; the protocol SHA-256 is
`576851e321034a0f6d3cb4d0818b3a5ca97975fcc6f2c7fa30b07ffd203a2d76`.

There are 72 fresh starting-position groups: 32 training, eight development and
32 confirmation. Both destinations remain together in their starting-position
group. Starts span x=0.287–0.311 m and y=−0.042–−0.018 m, with the configured
initial block height 0.082 m. This small interpolation domain is a deliberate
first proof, not general spatial manipulation. A 3 mm inset from the prior
fixed primitive's envelope buffers camera error; estimates are not clipped or
replaced with simulator truth.

| Registered stage | Budget | Role |
| --- | ---: | --- |
| Initial teacher diagnostic | 4 trajectories | Verify actual motion, initial camera acquisition and complete teacher placement |
| Training teacher demonstrations | 64 episodes | Two goals for each of 32 training starts |
| Development teacher demonstrations | 16 episodes | Prediction diagnostics; excluded from fitting and normalization |
| Mixed-control corrections | 32 episodes | One training-only round; teacher chosen with probability 0.5 |
| Model development | 64 rollouts | Initial/corrected × full/blind × 16 development cases |
| Frozen confirmation | 256 attempts | 64 cases each for learner, matched blind model, fixed primitive and teacher |
| Negative controls | 16 attempts | Blank initial RGB, invalid old initial timestamps, malformed extrinsics and forced-open gripper |

The maximum registered native workload is 452 attempts, with no replacement of
failed starts or hidden retry. A 20-second complete episode contains 501
320×240 RGB-D pairs at 25 Hz, 1,000 motor intervals at 50 Hz, and 20,000 native
physics ticks at 1 ms. Camera capture/availability timestamps are simulated
clocks; wall-clock processing is recorded separately. They do not establish
physical camera-to-actuator latency.

## Inputs, supervision and acceptance

The CPU NumPy model has two 64-unit hidden layers and predicts six absolute
joint targets plus one jaw target. Its 269 features contain an explicitly
historical initial visual anchor, destination, elapsed time, encoders, jaw
measurements, encoder-derived link6 body pose, previous applied servo references and
five causal history samples at 0, 40, 80, 120 and 160 ms. Standardized
features are clipped to ±10 at inference. Current missing object estimates have zero numerical
fills, an unavailable mask and their source frame's original age. The historical
anchor is never presented as a new live observation. The frozen recording names
the link6 body-origin pose `tcp_position_estimate_m` and
`tcp_quaternion_estimate_wxyz`; these legacy field names do not identify a
calibrated gripper contact point. The native task scorer instead uses a distal
site offset by 0.12003 m along link6’s local z axis. The independent auditor
reconstructs this explicit offset when comparing the two coordinate records.

Teacher phase, current teacher command, native object/contact truth and future
observations are excluded. Teacher commands are synthetic labels from a
privileged authored controller. In the mixed-control round, the one teacher
query per step is stored separately from the command actually executed; it is
not described as an operator correction or a proven recovery. The blind model
uses identical source episodes and initial context but has all live and
historical sensor feedback masked out. It retains past command history.

Each fit uses seed 17, 50 epochs, batch size 512 and Adam learning rate 0.003.
Normalization uses training data only, including explicit missing fills and
masks. Candidate selection uses development task success, critical contacts,
then prediction error normalized by the same physical action scales across
candidates. The matched blind model is taken from the selected stage.

Before training, every training start must have at least one successful usable
teacher demonstration, and each goal must have at least 28 successes. Final
qualification requires all of the following:

- At least 58/64 strict learned placements on the scheduled unseen cases.
- Positive 97.5% paired bootstrap intervals against both fixed and blind
  comparators, resampling whole starting-position groups with both goals paired.
- Positive comparisons also on at least 24 common admitted groups, preventing
  initial comparator refusals from being the only basis for improvement.
- No increase in the frequency of critical-contact episodes versus the fixed
  comparator; all negative controls behave as specified.
- Complete attempt accounting, independent native scoring, no privileged
  learner inputs and no teacher intervention during evaluation.

The teacher comparator has privileged simulation observations and is an upper
reference, not evidence of human-video learning. Installation alone cannot
award qualification. Failed candidates remain publishable research findings.

The fault definitions are deliberately narrow: `initial_missing` zeros the
initial RGB image while the physical object remains in the scene;
`initial_stale` sets capture and availability times to −0.2 s at episode time
zero; `altered_calibration` scales one rotation-matrix element by 1.1; and
`open_gripper` overrides issued jaw targets to 0.035 m. The first three exercise
input-contract refusal. They do not establish detection of a physically absent
object, valid-but-old frames or small valid rigid calibration errors.

## Completed diagnostic evidence

![Native MuJoCo RGB frame at 10.00 simulated seconds, development start 00 and destination 0. The authored teacher controls the arm; the frozen tracker reports no object position in this frame. This is diagnostic evidence, not a learned-policy result.](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/viewer/frames/development-00-goal0/0250.png)

An independent reader verified **24,040 files**, exact camera/control/native
cadences, source and frame bindings, issued/query separation, camera-position
errors and independently recomputed task scores. All four teacher diagnostics
completed strict placement successfully. They represent **two development
starting positions, each with two goals**, not four independent starting groups.

| Diagnostic | Initial camera error | Available-pose error p95 | Available decisions | Longest missing stretch |
| --- | ---: | ---: | ---: | ---: |
| Development 00, goal 0 | 0.196 mm | 0.314 mm | 620/1,000 | 7.24 s |
| Development 00, goal 1 | 0.196 mm | 0.504 mm | 618/1,000 | 7.28 s |
| Development 01, goal 0 | 0.249 mm | 0.816 mm | 602/1,000 | 7.96 s |
| Development 01, goal 1 | 0.249 mm | 0.816 mm | 604/1,000 | 7.92 s |

Native object displacement reached 138.6–160.0 mm, confirming actual
manipulation rather than a static-camera or replay-only check. Errors above
are conditional on the tracker returning an estimate. Adjacent 50 Hz decisions
often reuse the same 25 Hz frame and are not independent tracking trials.

The major diagnostic finding is **where visibility disappears**. An offline
analysis of sealed teacher-phase and perception sidecars found zero available
object estimates during close, lift, transfer, place and release: **0/749 motor
decisions** across the four trajectories. Approach and final verification were
fully visible. Teacher phase was used only to analyze this evidence, never as
a model input.

Therefore sub-millimetre available-pose errors must not be described as tracking
throughout manipulation. During the contact-heavy sequence the
learner must rely on measured robot state and its historical initial visual
context. This is a measured limitation of the frozen upright-block tracker.
The registered experiment tests those remaining signals without adding camera
views or substituting native object positions; its final results appear below.

[Inspect the real diagnostic camera frames and action traces](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/viewer/), [independent audit](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/viewer/diagnostic-audit.json), [phase coverage analysis](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/viewer/diagnostic-phase-coverage.json) and [registered protocol](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/viewer/protocol.json). The compact viewer contains diagnostic evidence only. It preserves every motor decision and copies native RGB frames byte-for-byte at 2.5 Hz from the original 25 Hz capture. Its display sampling is distinct from the camera cadence used by the controller. Full raw recordings remain separate from this compact diagnostic subset.

## Completed demonstration and data-quality evidence

All **80/80 registered teacher demonstrations** completed strict placement:
64/64 training episodes and 16/16 development episodes. The development
recordings remain excluded from fitting and normalization. All 32 training
starting-position groups have usable successful demonstrations with an
available, bound initial camera estimate. Each training destination has
**32 successful usable episodes**, exceeding the registered minimum of 28;
the sealed teacher-data gate passed with no missing training group.

| Teacher evidence | Strict successful / scheduled | Data use |
| --- | ---: | --- |
| Initial diagnostics | 4/4 | Initial camera and task checks; not additional fitting episodes |
| Training demonstrations | 64/64 | Issued-action supervision for 32 training starts × two goals |
| Development demonstrations | 16/16 | Held-out prediction and missing-observation diagnostics |

An independent readback passed with **504,840 files verified and 84 native task
scores recomputed** across the four diagnostics plus 80 demonstrations. This
count includes the earlier four diagnostics; it is not 84 additional episodes.
The audit reports no errors and executes no physics, estimator or learned
model. Its protocol, sealed demonstration receipt and stage-integrity identities
match the original registration and saved bytes.

The retained `demonstrations-audit.json` proof has SHA-256
`3405d789c0b79066d142d476f1b45468463d41516350d7c80766881a6cecf753`;
its evidence-inventory SHA-256 is
`319074d85fcd84758d3cc464e6040f2032832179d354c4adcb86f300b743f811`.
This is an audit of the diagnostic and demonstration stages only: no fitting
stage or student rollout is covered by this proof. Successful teacher actions
establish usable synthetic supervision, not successful imitation or transfer.

## Measured host context

The retained `hardware-native.json` describes a shared local **Apple M5** host
(`Mac17,4`, arm64) with **10 physical / 10 logical CPU cores and 24 GiB of
installed memory**. The environment is Darwin 25.5.0, Python 3.12.9, NumPy 2.5.3
and MuJoCo 3.15.0, with one native worker. Native physics and the policy use the
CPU; rendering uses the existing macOS CGL graphics path. No CUDA requirement
is introduced. The hardware-context file SHA-256 is
`fdc8007b27219c7c91dc600ae54a95f4a8abc64565db652002faa0f5a1e9f6d0`.

Installed memory is host capacity, not the model's measured memory requirement.
This shared-host setup does not establish minimum hardware, dedicated-host
throughput, other-platform compatibility or physical sensor-to-actuator
latency. Final pooled inference and full-loop wall-clock measurements appear
below and agree with independent full-study readback. Fit costs are separate from
the much longer rendering, recording and verified export.

## Initial learning result: prediction is not execution

The sealed initial fitting stage contains 64,000 training samples and 16,000
held-out teacher-observation samples. The full learner fitted in 6.56 seconds
and occupies 521,715 bytes; the matched blind learner fitted in 5.99 seconds
and occupies 517,296 bytes. Shared byte-verified data export took a further
685.44 seconds. These are measured offline wall-clock durations on the shared
host, not video-to-skill completion times. Process peak RSS reached 643,891,200
bytes after the full fit and 789,200,896 bytes after the blind fit; the second
value is a process high-water mark, not the blind model's isolated memory use.

Both initial candidates completed all 16 registered development attempts and
made **zero strict placements**. The full learner's held-out action-scale MSE
on teacher observations was 0.0001133, yet its closed-loop success was 0/16.
The blind learner also scored 0/16. One-step prediction on a demonstrated
state does not establish execution after the model's own actions change it.

## Corrected development: fewer contact violations, incomplete acquisition

The single registered mixed-control correction round and corrected fitting are
complete. Both development cohorts retain every scheduled attempt: 16 full and
16 blind attempts per candidate. All 64 learned trajectories and the 16 exact
matched teacher demonstrations are complete with valid native numeric scores.
No failed attempt was dropped, replaced or retried.

| Development cohort | Strict placement | Never commanded jaw below 20 mm | Native grasp / lift | Critical-contact attempts | Rate-limited intervals |
| --- | ---: | ---: | ---: | ---: | ---: |
| Matched teachers | 16/16 | 0/16 | 16/16; 16/16 | 0/16 | 1,200/16,000 |
| Initial full learner | 0/16 | 0/16 | 0/16; 0/16 | 16/16 | 15,706/16,000 |
| Initial blind learner | 0/16 | 2/16 | 2/16; 0/16 | 10/16 | 1,127/16,000 |
| Corrected full learner | 1/16 | 12/16 | 4/16; 4/16 | 0/16 | 187/16,000 |
| Corrected blind learner | 1/16 | 0/16 | 16/16; 16/16 | 9/16 | 1,175/16,000 |

The early-closure pattern from the first initial student recurs across all 16
initial full-learner attempts. Each first commands a single-finger jaw target
below 20 mm while the distal tool is more than 20 mm from the block in XY. The
median distance is **129.25 mm** and median command time **1.22 s**. None acquires
bilateral object contact. Matched teachers first command closure at median
**10.45 s** and **3.89 mm** XY separation. The initial full learner records
**25,368 critical-contact native ticks**, versus zero for these teachers.

The corrected full learner has a different observed failure pattern. It records
zero critical-contact ticks and far fewer rate-limited intervals, but **12/16
attempts never issue a jaw target below 20 mm and never acquire bilateral or
strict grasp**. The other four first command closure at median **15.18 s** and
**5.40 mm** XY separation; all four reach the native grasp and lift predicates,
and one completes strict placement. The corrected model therefore has lower
observed contact violations without reliable task completion. These contact
counts do not establish hardware safety or physical transfer.

The corrected blind learner reaches grasp and lift in all 16 cases while
completing just one strict placement. Nine attempts contain critical contacts,
totalling **2,816 native ticks**. Grasp acquisition alone is not the full task.
This compact analysis does not attribute those remaining failures to retention,
release, settling or retreat without their corresponding task predicates.

Here “closure” means an **issued single-finger target**, not measured finger
closure or total jaw aperture. Native geometry comes from the preceding 1 ms
sample at the command boundary. The two 20 mm thresholds are descriptive
posthoc choices, not new acceptance gates. Closure-time and distance medians
include only attempts with a closure command: four corrected-full attempts and
14 initial-blind attempts, with the no-command cases kept in all scheduled
denominators. No camera/media audit or new model inference was performed by
this analysis.

## Frozen choice and remaining evidence

The registered ranking selects the **corrected full learner and its matched
corrected blind baseline**. The full learner improves from 0 to 1 development
placement and from 25,368 to zero critical-contact ticks. Its teacher-state
action-scale MSE increases from **0.0001133175** to **0.0001734760**. Success has
priority over contact counts, which have priority over prediction loss; the
saved choice matches that ordering. Lower teacher-state prediction error did
not correspond to better task completion in these two candidates.

A separate read-only check reconstructed the ranking and verified all 11
model/stage artifacts named by the freeze, the fixed comparator and protocol.
The frozen-selection file SHA-256 is
`88ef3ae4f5639dc00c2570debf099dc33d52445c9600d3d681aa634c4ddb4087`.
The saved pre-confirmation flag is reported; this compact check does not
independently establish temporal chronology.

[Read the descriptive findings](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/development/FINDINGS.md),
[all development metrics and source bindings](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/development/development-failure-summary.json),
and the [selection cross-check](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/development/selection-check.json).
The numeric summary checked 581 source-file identities and all 19 frozen
runtime hashes; its SHA-256 is
`ae3790c8afcbbe6740ae7b0c736798a871c10a89d6eb80e5773ca0a0c158ea3d`.
This remains **development-only analysis**. The subsequent full-study audit
checks the sealed development and confirmation records independently; these
descriptive findings do not replace frozen confirmation or award qualification.

The observed change is consistent with several possible mechanisms, including
closed-loop distribution shift and action-target scaling, but this study has
not isolated their causal effects. No hyperparameter change, extra correction
round, confirmation-driven selection or hidden retry has been introduced.
The native renderer also reports a depth-accuracy limitation when
`ARB_clip_control` is unavailable for `mjDEPTH_ZEROFAR`. Repeated warnings are
not independent accuracy trials, and their presence does not establish that
they caused these failures. Synthetic camera transforms do not validate real
depth measurements.

## Frozen confirmation and qualification

All **256 registered confirmation attempts** completed: 64 per controller over
32 unseen starting-position groups, with both goals paired. There were no
initial refusals, missing outcomes or replacement scenes. All 32 groups were
also common-admitted for both learned-versus-comparator comparisons.

| Controller | Strict success | Success rate | Native grasp / lift / carry | Critical-contact attempts / ticks |
| --- | ---: | ---: | ---: | ---: |
| Selected full learner | 5/64 | 7.81% | 22 / 21 / 6 | 2/64; 122 |
| Matched blind learner | 9/64 | 14.06% | 63 / 63 / 38 | 34/64; 14,953 |
| Fixed primitive | 31/64 | 48.44% | 64 / 64 / 54 | 30/64; 2,306 |
| Authored teacher | 64/64 | 100% | 64 / 64 / 64 | 0/64; 0 |

The paired bootstrap uses the registered 10,000 resamples, seed 311402 and whole
starting-position groups. The **97.5% intervals** for learner-minus-comparator
strict success are:

| Comparison | Difference | Paired interval, percentage points |
| --- | ---: | ---: |
| Learner minus fixed | −40.625 percentage points | [−60.9375, −20.3125] |
| Learner minus blind | −6.25 percentage points | [−20.3125, +7.8125] |

The common-admitted intervals are identical because no group is excluded.
Neither comparison is positive. The learner also misses the required 58/64
placements. The no-contact-increase gate passes: 2/64 learner attempts have
critical contacts versus 30/64 fixed attempts. Complete accounting and the
recorded no-privileged-input/no-teacher-intervention gate pass. The overall
result is **`qualified: false`, `research_candidate`**. The success and
superiority gates are retained unchanged; the fixed comparator scored 31/64,
so this result is not explained by a perfect-comparator ceiling.

All **16 negative controls** behave as registered: the 12 blank-initial-image,
invalid-timestamp or malformed-extrinsics attempts refuse motion; the four
forced-open-gripper attempts complete their simulation budget and fail the
independent task scorer. Their scope remains the narrow input-contract and
gripper interventions defined above. Refusals do not become successful tasks.

## Where completion breaks

The selected learner fails to reach the native grasp predicate in **42/64**
confirmation attempts. Of the 22 that grasp, one fails to lift. Of the 21 that
lift, **15 fail supported carry**; none of those 21 records retention loss under
the frozen predicate. Five of the six supported-carry attempts meet the
terminal composite and strict placement. Among those six, one fails each of the
posthoc release, settling and retreat dwell checks. These are conditional,
overlapping component counts, not extra attempts or new qualification gates.

Terminal appearance alone would overstate the fixed controller: it meets the
terminal-only predicate in **64/64** cases, while strict success is **31/64**.
Supported carry succeeds in 54/64, retention is lost in 17/64 lifted cases, and
30/64 contain critical contacts. These failures overlap and must not be summed.
The strict path checks expose failures that a final-image match would miss.

The blind learner's 63 grasps and 38 carries also exceed the full learner's 22
and six, yet neither candidate reliably completes the task. This bounded
comparison does not show that feedback is generally harmful. It shows that
this learned use of the available feedback, action representation and training
data has not delivered reliable closed-loop control. Causal ablations remain
future, separately registered experiments.

The full learner has an available position estimate on **55,694/64,000 motor
decisions (87.02%)**, with a longest missing interval of **10.68 s**. The blind
controller's recorded tracker is available on 34,884/64,000 decisions, although
its sensor features are masked. Coverage depends on the visited trajectory:
staying away from manipulation can keep the object visible. Higher aggregate
coverage is therefore not evidence of better manipulation or tracking during
contact.

## Local computation and timing

The selected corrected full model occupies **543,977 bytes** and fitted/saved
in **10.95 s**; its matched blind model occupies **540,371 bytes** and
fitted/saved in **13.69 s**. Both use the same 96 training episodes: 64 teacher
demonstrations plus the single 32-episode mixed-control round. Shared verified
export took **471.30 s**, counted once for the pair. The process peak RSS for
that stage is **1,055,539,200 bytes**, including export and preceding fits; it
is not an isolated inference memory requirement. Mixed-control episodes are
retained even when unsuccessful; only 1/32 completed strict placement.

| Confirmation controller | Policy computation p50 / p95 | Full boundary p50 / p95 |
| --- | ---: | ---: |
| Full learner | 0.178 / 0.447 ms | 18.775 / 36.432 ms |
| Blind learner | 0.095 / 0.254 ms | 17.821 / 35.627 ms |
| Fixed primitive | 0.166 / 0.439 ms | 18.158 / 37.593 ms |
| Authored teacher | Not separately reported | 18.587 / 38.634 ms |

Each row pools **64,000 control-boundary samples**; these are not averages of
episode percentiles. The full learner's tracker computation p95 is **13.393
ms**, pooled over 32,064 camera acquisitions. Its full-boundary timing exceeds
20 ms on **31,708/64,000** samples. The simulator pauses for host computation;
sub-millisecond policy inference does not establish a 50 Hz physical control
deadline or real camera-to-actuator latency. The shared Apple M5 host context
above remains a measured configuration, not a minimum hardware specification.

## Final evidence and independent verification

All nine registered stages are sealed. The saved study report is complete and
the posthoc summary covers all 452 scheduled attempts, including the 12
intentional negative-control refusals. It preserves failures and reconstructs
terminal component predicates from native numeric evidence without rerunning
physics or changing the model.

[Complete saved report](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/final/report.json),
[posthoc tables](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/final/summary.md)
and [machine-readable summary](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/final/summary.json)
are retained with their byte identities. Report SHA-256:
`0234118aea1d67ac0e2747fa5ab401e862654cb5396b2dddb0826511bf09bcbd`.
Summary SHA-256:
`569a967597e6e0e94a7e9cfbda3837bce568dd250c695e82d25b7a9a1e3049dd`.

## Fresh installation and execution

A clean local installation of reviewed source
`ab9082e1ee2b2cc91c7c5fd3f1358727b3789c66` executed the selected corrected model
outside the checkout. Independent readback verified **5,010 files, 20,000
native ticks and 1,000 control intervals**; the complete capture contains 501
RGB-D pairs. All 19 registered runtime files matched the freeze, and the
installed model SHA-256 was
`d176b391b61b5e331bcfe1676107c53f4c0ae6110414dfed4b07d57cbb883b0e`.

The single run completed with zero teacher queries, but **failed the independent
task scorer**; the CLI returned exit code 1. This verifies honest installed
execution and failure reporting, not task qualification. It was one
out-of-study attempt with **zero retries**, excluded from the 452 registered
attempts and the 64-case learner denominator. A fresh local Training Studio
check installed the same corrected model; installation did not change its
research-candidate status.

[Public installed execution proof](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/final/installed-smoke-audit.json)
SHA-256: `ed0b411aaa1bafe94a6411851a42adcb4695b927801a02967fb2aedb21c1811a`. This public derivative replaces only two private
local paths; all numeric evidence and outcomes are unchanged. It binds the
retained original proof SHA-256
`770622dfd7b30c7f71cf762a35d05386d50598ec1bec9bdd8db5a0b1b837ee25`.

The independent full-study audit completed on **9 November 2026** with **zero
errors**, verifying **2,384,532 files** and reconstructing all **452 registered
attempts and saved native scores**. This count includes the 12 early refusals;
it does not mean 452 complete twenty-second executions. The remaining 440
recordings contain the complete execution interval. The auditor checks source
bindings, group splits, clocks, action authority, model selection and final
gates from sealed bytes and numeric traces. It executes no physics, estimator
or learned model. All 19 registered runtime file hashes remain unchanged.

[Read the full-study audit](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/final/full-study-audit.json)
SHA-256: `2774959f8e1a5be7283cea129da0ffea000d8375eb445140c8c7a60ab15efed8`.
This passing consistency audit leaves the model **unqualified**. No acceptance
threshold, model selection or failed scene was changed after confirmation.

[Inspect the first registered learned attempt](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/viewer-v02/).
The new reader includes the four diagnostic recordings and the first student
entry in protocol order, selected without looking at success. It retains all
motor decisions and samples existing RGB frames for display at 2.5 Hz. The
original diagnostic viewer is unchanged. This five-recording display is an
inspectable subset, not the full 452-attempt archive or an aggregate success
claim.

**The GitHub software preview is public and anonymously verified.**
[Download the preview release](https://github.com/Dvidia-Inference/dvidia-training/releases/tag/camera-to-action-v0.1-preview)
for the wheel and verified audit summaries. Publication of the approximately
36.28 GB frozen dataset archive on Hugging Face remains incomplete; this
software preview does not supply that full archive or qualify the skill.

This study supplies a reproducible camera-and-action learning loop whose first
bounded synthetic task candidate fails qualification. It cannot establish arbitrary video-to-skill transfer,
arbitrary-arm compatibility, real-time hardware operation, real-world depth or
contact calibration, shoe manipulation or physical qualification. Subsequent
preregistered experiments should isolate action scaling, rate-consistent target
representations and correction-data collection. Separate perception experiments
can measure a second camera or another explicitly uncertain observation method
during grasp and carry, where the present tracker often provides no object
measurements. None of those interventions was tested in this frozen study.

## Engineering follow-up — 9 November 2026

These bounded engineering checks prepare the next experiment without changing
this study's 452 attempts, selected model or 5/64 result.

An optional [verified derived-data cache](https://github.com/Dvidia-Inference/dvidia-training/blob/cc6746a95669cfd69a49e32a3e6c2e5a42ce2546/docs/camera-action-export-cache-v01.md)
reuses strictly exported features and action labels with independently retained
source, protocol and manifest pins. For one sealed demonstration containing
1,000 samples, strict export took **1.635732 s**; the median of three verified
cache reads was **2.952292 ms**, approximately **554× faster for this preparation
step**. The three reads measured 2.952292, 2.830833 and 3.214208 ms. Numeric values and the dataset identity matched the strict export.
This is a single-episode, warm local preparation measurement, not a training,
simulation, end-to-end video-processing or robot-success improvement. Cache
creation, source validation and data collection remain separate costs. Cache
hits validate the pinned derived arrays; they do not reverify all raw images
and sidecars. [Retained measurements and prior attempts](https://github.com/Dvidia-Inference/dvidia-training/blob/cc6746a95669cfd69a49e32a3e6c2e5a42ce2546/benchmarks/camera-action-export-cache-v01/README.md)
disclose a rerun after a Windows file-cleanup portability fix. These shared-host
runs are not a controlled before/after comparison of implementation speed.

A [complete training-data follow-up](https://github.com/Dvidia-Inference/dvidia-training/blob/77624ca66ace1b3aa81d076fefc294fca9015711/benchmarks/camera-action-export-cache-v01/README.md#complete-training-dataset-follow-up)
then checked all **64 teacher and 32 mixed-control training recordings**:
96,000 samples across 32 starting-position groups. The two caches took
**160.671 s to construct once**, including strict source validation. Three
subsequent validated loads of both caches plus ordered combination took
**0.347, 0.280 and 0.279 s**. Separate blind projections and comparisons with the
four original sealed model dataset/provenance identities took **0.596, 0.491
and 0.490 s**; all **12/12 identity checks passed**. The full training source
inventory contains 577,056 files / 18,971,932,592 bytes; derived caches total
108,804,852 bytes. Whole-process peak RSS was 1,631,633,408 bytes.

The [reviewed full-training receipt](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/final/training-cache-validation.json)
binds the four model hashes, source and protocol, original executed script and
all measurements. This was one shared-host offline preparation run with no
matched strict-export baseline, so it establishes no whole-pipeline speedup
ratio. The roughly 0.28 s repeated load-and-combine measurement excludes model
fitting and the separately reported identity checks. No development or
confirmation recordings were exported, no policy or physics ran, and the
learned **5/64 result remains unqualified**. The [portable reproduction helper](https://github.com/Dvidia-Inference/dvidia-training/blob/77624ca66ace1b3aa81d076fefc294fca9015711/benchmarks/camera-action-export-cache-v01/validate_full_training.py)
changes only host paths into required CLI inputs; its [provenance](https://github.com/Dvidia-Inference/dvidia-training/blob/77624ca66ace1b3aa81d076fefc294fca9015711/benchmarks/camera-action-export-cache-v01/full-training-script-provenance.json)
distinguishes it from the measured original. This benchmark-only follow-up
leaves the released `cc6746a` code, dev5 wheel and preview release unchanged.

An [action-reference preflight](https://github.com/Dvidia-Inference/dvidia-training/blob/cc6746a95669cfd69a49e32a3e6c2e5a42ce2546/docs/camera-action-v02-preflight.md)
checks whether alternative target representations preserve the existing
actuator reference trajectory. All **4,242 bounded numerical cases**, each
containing **20 microsteps**, agree within the predeclared **1e−12 native-unit
absolute tolerance**; all **24 negative contract/input cases** are refused.
Closed-form floating-point projection is not universally bitwise identical to
the original repeated updates. This check establishes numerical equivalence
under its declared tolerance, not simulated task success or physical actuator
accuracy. It reads no v0.1 recordings and trains no policy.

The cache runtime and numerical preflight have [separate reviewed source in PR 33](https://github.com/Dvidia-Inference/dvidia-training/pull/33),
pinned at `cc6746a95669cfd69a49e32a3e6c2e5a42ce2546`, and remain outside the
frozen v0.1 runtime. The [cache preview release](https://github.com/Dvidia-Inference/dvidia-training/releases/tag/camera-action-cache-v0.1-preview)
is public as a non-latest prerelease. Anonymous downloads verified all six
assets against their recorded SHA-256 identities: the dev5 wheel, preparation
benchmark, public CLI proof, numerical protocol, numerical report and checksum
list. Fresh offline macOS wheel installation also passed.

All 36 CI checks for the released `cc6746a` head passed. In that pull-request workflow, the six
Visual jobs—`ubuntu-latest`, `macos-latest` and `windows-latest`, each with
Python 3.12 and 3.13—passed both pinned-cache verification and the existing
footage, visual-training and local-studio tests.
[Public release and CI verification](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/final/cache-preview-release.json)
retains the exact tag, commit, asset hashes and individual job links. These
software checks do not establish reproduction of the frozen native study on
all platforms or physical robot qualification.

The next control comparison also needs the exact preceding servo reference.
The existing normalized float32 features cannot reconstruct that float64
quantity. A separate training-only export now preserves each episode's reset
reference and preceding **actually applied** reference, together with the next
feasible reference toward its teacher label. In mixed-control episodes that
label is a teacher suggestion; it is not represented as the executed action.

One [metadata-only validation](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/final/training-reference-validation.json) covered all **96,000 training samples** and
preserved the original parent and combined dataset identities. Construction
took **10.798 s** and produced 10,832,870 bytes. Three verified paired loads took
**0.303, 0.258 and 0.233 s**; combination and dataset identity checks are timed
separately. The validator read 294 registered metadata files. It checked final
reference endpoints against the recorded commands, inheriting raw verification
from the pinned parent caches; it did not reread images or native microstep
traces, train a model, or run physics. Whole-process peak RSS was
1,238,597,632 bytes on a shared macOS arm64 host with Python 3.12.9 and NumPy 2.5.3.

A separate synthetic initialization check uses identical network weights and
a common output scale for absolute and reference-relative predictions. All
**64 reset-context comparisons were bitwise equal before clipping**. Later
predictions differ by the preceding-reference offset, as intended. These are
synthetic input vectors, not real reset observations or learned task results.
Together, [the reviewed source in PR 35](https://github.com/Dvidia-Inference/dvidia-training/pull/35)
prepares a fair control-representation experiment. It does not establish that
small corrections improve grasping; that requires the separately registered
training and closed-loop comparison. The frozen 5/64 result is unchanged.
The [exact source and reproduction instructions](https://github.com/Dvidia-Inference/dvidia-training/blob/f2676d3176f6a8cd1a05833d461ba40562f283f1/benchmarks/camera-action-reference-export-v01/README.md)
separate the original measured helper from its portable version.

This is an engineering-tool preview, not a qualified skill. A new registered
learning experiment must measure whether the representation changes improve
grasp timing, transport and complete placement.

A [separate model follow-up](https://research.dvidia.org/papers/reference-action-v01/) now reports
two real CPU fits and all eight native development attempts. Both candidates
complete zero of four strict placements; neither is promoted. It retains the
earlier infrastructure refusals and leaves this frozen study unchanged.

## Source and reuse

[The reviewed delivery source](https://github.com/Dvidia-Inference/dvidia-training/tree/d89b24c82b7534a439bbaa3cac3ef5c9d3b80088)
is frozen at `d89b24c82b7534a439bbaa3cac3ef5c9d3b80088`, separately from the
unchanged runtime freeze above. It contains the capture, mask-aware exporter,
CPU model, study runner, installer and local Training Studio flow. The exact
source archive and wheel retain all 19 registered runtime identities.

This delivery adds only deterministic ordering of publication-card gate rows
and its regression test. It changes no simulation, learning, installer or UI
bytes from the `ab9082e` delivery used for the native execution above. Fresh
offline installation, 12 CLI checks, four renderer checks and six package
resources passed; the six resources also match the previous delivery.
[Public delivery verification](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/final/delivery-check.json)
binds source archive SHA-256
`a9d95981d887e9edfa461c780efd07711e4660170d52ea7571b898a517a232dc`
and wheel SHA-256
`b6b116156028b8c7e2d2c3d6a061f75d7f26e5671e59639569dd766d51af2eb3`.
These packaging checks execute no policy or physics and leave the original
installed-run attribution intact.

The dataset card and four model cards were generated from the same sealed
report and verified full-study audit. Their [card receipt](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/final/card-receipt.json)
has SHA-256 `b4b976be7f33e0d40e3aead02a30bdfed32a3c0491bfbf7aee7afde1f4b73fa3`.
It binds the initial and corrected full/blind model files, recorded fit costs,
failed qualification gates and source identities. All cards retain synthetic,
simulation-only research-candidate labels. The five cards render identically
from saved, inspected and reversed gate ordering; no evidence or qualification
value changed to resolve the publication mismatch.

The [GitHub preview release](https://github.com/Dvidia-Inference/dvidia-training/releases/tag/camera-to-action-v0.1-preview)
is a public, non-latest prerelease at the exact delivery source above. Its five
assets are the installation wheel, full-study audit, JSON summary, Markdown
summary and checksum list. Anonymous downloads matched their recorded byte
identities; the [release verification receipt](https://research.dvidia.org/downloads/topics/physical-grounding/runs/2026-11-01-camera-to-action-v01/final/github-preview-release.json)
binds the tag, source commit and all five asset hashes.

The intended full dataset destination remains
[Hugging Face · Dvidia/piper-camera-action-v01](https://huggingface.co/datasets/Dvidia/piper-camera-action-v01).
The approximately 36.28 GB frozen archive has **not completed publication and
anonymous revision-pinned readback**. The GitHub assets do not include that
archive, raw recordings or the original model bundle. No download here is
labeled a qualified skill.

Original DVIDIA code and generated records use MIT; PiPER and simulator notices retain their upstream terms. The viewer includes actual native renderer output and no private human footage, credentials, model weights or redistributed robot meshes. Research writing follows this site's CC BY 4.0 license.