Research / Embodied AI safety

SENTINEL-EVC / ROBOTICS

Sentinel-EVC

Check the action. Trace the execution.

From the final action to traceable execution.

v0.3 · Open-source execution verification workbench · Updated 03 October 2026

Sentinel-EVC connects final-action verification, one-use authorization and inspectable execution. Core research and implementation are substantially complete, with open-source maintenance continuing. The latest update adds incremental rechecking, Launch Gate integration diagnosis and writer-entry hardening.

Original conceptual artwork of a robot, an action trajectory and a validation boundary
Original concept artwork · Recorded experiments and trajectories below

Why must a check keep up with change?

A policy’s action may be retimed, repaired, spliced or transformed between frames. A verdict about the original plan may not cover the motion finally sent out. Sentinel-EVC places the check after transformation and before dispatch, so the decision concerns the version that would execute.

The object of verification is the final action handed to the executor.

Connect the check, the permit and the evidence.

The system asks three questions: does the action pass the current checks, does its permit still match the context, and can an independent reader reconstruct what happened? A changed plan or context returns to rechecking; unsupported candidates remain unknown.

  1. 01

    Check the final version

    Inspect the transformed plan against physical constraints and its bound consequence profile.

  2. 02

    Limit permission in time and use

    Bind a permit to the plan, context and generation, then recheck fresh feedback at submission.

  3. 03

    Retain an inspectable trace

    Distinguish submitted, accepted and observed actions, then replay and verify independently.

Execution chain and recovery rules
Mechanism diagram: upstream action, transform, final checks, expiring one-use permit, single writer, and signed evidence.
Conceptual mechanism diagram; the capability matrix states each profile's actual runtime integration.
  1. 01

    Propose

    A VLA, planner, or recorded policy proposes an action block.

  2. 02

    Transform

    Retiming, repair, and frame conversion produce the plan that would actually execute.

  3. 03

    Recheck

    Physical checks and the bound consequence-prediction profile evaluate the final candidate; unsupported actions remain unknown.

  4. 04

    Permit

    The permit binds plan and context, with a deadline and one use.

  5. 05

    Execute

    The single writer rechecks fresh state and distinguishes submitted, accepted, and observed commands.

  6. 06

    Verify

    Events, results, and asset digests enter signed records for offline replay and verification.

Revocation prevents further old-generation submissions. Recovery confirms cancel and drain, obtains post-revoke feedback, and resumes with the current epoch and explicit approval. Actions already accepted by the controller remain part of the record.

Recheck what changed. Preserve what still holds.

When a robot revises its next action chunk, the earlier check need not be discarded in full. Sentinel records which parts remain valid, checks the affected motion again, and binds the approved action bytes to the software writer. This connects the cost of repeated verification with the question of what actually gets sent.

Four candidates · 0.002 rad changes
1.619×Full / incremental median time; 95% root-cluster CI 1.276–1.755.
Four candidates · 0.010 rad changes
1.092×Full / incremental median time; 95% root-cluster CI 1.035–1.253.
Unique candidate inputs across ten cells
249,120747,360 paired decision comparisons include three technical repeats and warmup.
Incremental / full decision differences
0Across the retained grid, within the fixed static bare-arm geometry profile.

The four-candidate extension measures validation wall time on a Xeon Platinum 8474C CPU. It includes root-certificate amortization and the copying, hashing and issuance of the selected certificate: full versus incremental medians are 1.855 versus 1.145 ms for smaller changes, and 1.998 versus 1.830 ms for larger changes. The RTX 4090 D host GPU renders the recorded scene; it does not perform these geometry timings.

Original measured UR5e speedups and root-cluster confidence intervals across all ten configurations.
Original experiment figure. Each cell has 12 independent roots and three technical timing repeats. The four-candidate extension was frozen separately. Root costs are amortized over 40 measured cycles; intervals resample whole roots, not individual process repeats.
The retained late_suffix/root0 trajectory, rendered with the official UR5e visual model. The approved parent has no contact in 300 frames; the rejected revision has contact in 98. This is counterfactual kinematic replay explaining the rejection, not robot execution of the blocked motion.
Where reuse helps, and where it does not

One-candidate checking is a retained negative result: full / incremental time is about 0.749×, so incremental checking takes approximately 33.5% longer. At 128 candidates, incremental checking still misses the 50 ms budget in 235/1,440 and 344/1,440 measured cycles for the two change magnitudes. Validation gains therefore depend on reuse and amortization; they are not whole-loop latency or hardware-throughput claims.

The shared numerical CLI/App lifecycle now reuses exact remaining suffixes. In its fixed seed7 comparison, both versions submit, accept and observe 40 steps; the updated path creates one root certificate, makes no full suffix rechecks, and reuses 220 segments. This numerical product integration and the UR5e checker are separate experiments. Their combination with a real VLA executor remains a further integration study.

Check the integration that is actually installed.

An adapter can remain internally valid while changing which camera reaches the policy, the order or units of state fields, or the selected action cursor. Launch Gate compares those captured inputs and exact final action bytes with a reviewed reference on fixed probes. It locates the first changed field, blocks the trusted software writer, and exports a compact signed decision for independent inspection.

Signed-record adapter replay
10/10 BLOCKInjected integration configurations; 20 requests from ten task windows.
Independent GPU launch-probe study
6/6 BLOCKFault configurations; ten LIBERO tasks at state 46, with 40 fresh inference paths.
Blocked-case downstream software calls
0Every GPU fault also has zero writer-entry acknowledgements; env.step calls are zero.
Signed capsules independently reproduced
13 + 7Thirteen recorded and seven GPU-derived decisions; CPU reproduction retains their original verdicts.
Conceptual diagram separating pre-launch reference probes from final-request authorization and the live writer-entry check.
Explanatory concept diagram. A reference-probe PASS covers those finite inputs; the actual writer still requires final-request authorization and a fresh entry check. Signed evidence authenticates retained records under the selected key.

The GPU study runs the official SmolVLA preprocessor, model and postprocessor on an RTX 4090 D. Reference, unchanged, camera-swap and state-swap paths are inferred afresh for each task; the other faults are applied at their declared adapter stages to the official output. The recorded study instead remaps independently verified stored actions. Neither result turns CPU evidence reproduction into a new GPU inference run, and the no-motion launch study does not measure task success.

Diagnosis, controls and portable evidence

The published faults cover camera mapping, state order and units, output order, direction, gripper sign and chunk cursor. Repair hints show observed paired-field relationships rather than an established cause; Sentinel does not automatically apply a guessed correction. Unchanged integrations pass, behavior-preserving identity changes pass in the recorded suite, and different source inputs return REVIEW.

The GPU camera-swap capsule is 26,172 bytes before compression. It contains paired probe records, the decision, events and signed manifest, retaining camera commitments rather than raw pixels. A CPU reader can authenticate it and recompute the authoritative decision without model weights. Reproduction PASS means the original decision was reproduced—even when that decision is BLOCK.

For 200 CPU measurements of the unchanged recorded case, qualification averages 1.436 ms with P95 1.484 ms. This excludes inference, signing, network transfer and environment stepping. The packaged macOS App can import captured probes, show the first difference and export the signed reproducer; that App check reuses the probes and does not rerun the GPU model.

Installed KineGrant and RLSOK both accept 40/40 reference and camera-swap requests when each is newly and explicitly authorized; those accepts are intended behavior. RLSOK also detects declared release changes and invalidates the earlier approval. Sentinel adds the delivered workflow that captures observed VLA fields, aligns a reviewed reference, diagnoses the first mismatch and exports a signed offline decision. Custom adapters can add equivalent checks to neighboring systems.

Launch qualification covers captured integration behavior on finite probes and the trusted software-writer boundary. Collision clearance, sensor authenticity, physical stopping and hardware safety require their own validation.

Keep approval intact through the final handoff.

Initial admission and actual writing happen at different moments. A caller may modify its action buffer while the writer is preparing; feedback, context or time may also change. Sentinel now owns an immutable approved copy and rechecks live conditions at the writer-entry acknowledgement. The record distinguishes preparing a call, entering the writer and dispatching the final bytes.

Invalid dispatches · preceding → updated revision
40/40 → 0/40Four concurrent preparation conditions, ten probes per condition per revision.
Ordinary authorization faults denied
60/60Sentinel, KineGrant and RLSOK each; a shared capability, not a superiority claim.
Valid requests executed in the common grid
10/10Each of the three application authorization systems admits the valid control requests.
Authenticated, encrypted CycloneDDS delivery
70/70Byte-identical, ordered records between separate processes on loopback; software probes follow.

The updated path still invokes the preparation wrapper in all 40 probes. It dispatches the approved copy in ten caller-buffer cases and denies the other 30 cases when feedback, context or the deadline has become stale. Zero invalid writes therefore means zero invalid downstream dispatches; it does not mean that no wrapper ran or that all requests were rejected.

Inspect the preparation conditions and transport controls
Recorded concurrent preparation probes · invalid downstream dispatches
Preparation changePreceding revisionUpdated revisionUpdated outcome
Caller changes its buffer10/100/1010 approved copies dispatched
New feedback arrives10/100/1010 denied before entry
Execution context changes10/100/1010 denied before entry
Permit expires during preparation10/100/1010 denied before entry

These are actual thread-and-barrier comparisons between Sentinel revisions: 80 rows in total. Deadline probes wait on the real monotonic clock; ordinary authorization expiry controls use injected clock values and are reported separately. Exploratory KineGrant and RLSOK callbacks have different entry contracts and are not used for this ranking.

CycloneDDS 11.0.1 provides participant authentication, access control and encrypted delivery. The qualified run delivers 70/70 records byte-for-byte in order; forbidden-topic and no-ACL controls are denied, and an untrusted-CA peer delivers none. The earlier history-depth-one run delivered only 3/70 and timed out; that failure is retained. Application checks then use the received records with local state and clock fixtures.

The ordinary 70-record grid contains ten valid requests and 60 invalid attempts. KineGrant 2.65.5, RLSOK 1.5.12 and Sentinel all execute the ten valid requests and block the 60 faults. Transport security and the schema/HMAC reference lanes answer different questions. Sentinel's product contribution is the native VLA integration, live entry acknowledgement, separate execution cursors and signed task-specific replay.

These are recorded-action software-writer experiments, not robot hardware motion. They add no new policy inference or current-gateway latency benchmark. The earlier 0.802 ms gateway mean belongs to the preceding revision. Signed logs authenticate records; trusted gateway and writer adapters remain part of the boundary.

Authorize the final request. Inspect the execution.

Historical native-v3 engineering grid, retained before the buffer-copy and live-entry hardening. Its task outcomes and timings were not rerun on the current revision.

After official policy postprocessing, Sentinel binds the exact seven-value action bytes to feedback, context, environment step, queue revision and revocation generation. A short-lived, one-use permit is consumed immediately before the captured environment writer is entered. Submitted, accepted and observed actions stay distinct.

Original mechanism diagram: official policy, exact request, one-use permit, native writer and observed result.
Original mechanism diagram: official policy, exact request, one-use permit, native writer and observed result.
Task successes in each lane
33/40Direct = active · 7 failures retained
Identical final-action bytes
5,420All 40 paired action sequences
Invalid authorizations blocked
60/606 synthetic classes · 0 forbidden writes
Historical native-v3 admission cost
0.802 msP95 1.040 ms · 5,420 actions

Both lanes ran actual official SmolVLA inference on ten LIBERO-Spatial tasks and four initial states. All 40 episode pairs completed; seven reached the 280-step cap in each lane. The comparison measures integration compatibility, with four previously used diagnostic cells disclosed. The earlier 93/100 observation-only study uses a separate grid.

Frozen comparison, fault classes and timing scope
Original result figure: 40 paired cells, six authorization-fault classes and per-action cost.
Original result figure: 40 paired cells, six authorization-fault classes and per-action cost.

MuJoCo 3.3.7 · RTX 4090 D (24 GB) · initial states 46–49 · predict 50 / execute 10. Each lane succeeds in 82.5% of its 40 episodes; the 95% Wilson interval is 68.05–91.25%. All ten task denominators and seven failures remain in the downloadable records.

Action replacement, changed context, permit expiry, permit reuse, changed feedback and revoked generation were each tested ten times. All 60 invalid attempts were denied before the monitored writer. The fault lane also completed 8/10 tasks; those episodes have a separate denominator.

Authorization and admission timing ends before the writer. Inference, simulation, rendering and evidence serialization are excluded. This profile checks in-process request integrity; Panda collision, dynamics, WorldGuard and physical stopping remain separate, unvalidated requirements.

Physical robot interfaces remain read-only. A revocation prevents new admissions; an already-entered synchronous simulator call may finish.

Matching actions. One changed camera observation.

The new execution comparison aligns two independently verified runs across actions, feedback, runtime cursors, outcomes and authorization. On the same 40 pairs, it locates one camera-2 fingerprint change at task 0, state 48, step 4. The preceding feedback and current action agree; the cause of the camera difference remains open.

Final actions
0 / 5,420
Task outcomes
0 / 40
Runtime cursors
0
Camera-feedback frames
1 / 5,460

Different records / compared items · The intentional change from direct submission to active authorization is tracked separately.

Follow the difference back to its source.

The report links the changed field, execution step and both source frames. Action and fingerprint data are stored at step 4, while the preceding saved body pose is step 0. The App labels this gap instead of presenting a reconstructed pose as a recorded measurement.

Downloadable derived reports can be recomputed. Category gates pass for actions, cursors, outcomes and unexpected authorization differences, while the observed feedback difference remains a failing feedback gate. Changed report summaries are rejected.

Bring the research into an inspectable workbench.

The CLI and locally built macOS App share a local engine. Import a native run with an independently selected public key, choose from all 40 formal episodes, inspect actual task geometry and seven action components, then export and reimport the signed record. Browser replay and SSH experiments provide additional entrypoints.

Actual macOS App: selected-key import of the 40-episode native record, actual Panda task geometry and execution evidence.
Actual macOS App: selected-key import of the 40-episode native record, actual Panda task geometry and execution evidence.

Local acceptance records retain export and independent reimport verification. Distribution signing and notarization remain pending; signatures establish record integrity under the selected key, rather than sensor truth or physical execution.

Current modules and walkthrough
Final-action verification

Immutable plan digests, full physical checks, and bounded prefix reuse; mismatched bindings trigger full fallback.

Product core

Consequence prediction

Evaluate four sibling candidates, retain prediction envelopes and model/calibration identities, then select deterministically from the allowed set.

Numeric core; GPU models are separate research

Permits and runtime state

Time-boxed single-use permits, fresh feedback, generation revocation, confirmed cancel, and drain-based recovery.

Product core

Signed evidence and replay

Retain run, model, prediction, and event records; verify independently, filter events, and download ZIP evidence.

Local workbench

3D asset inspection

Import OBJ, STL, MJCF, and URDF; preview parsed geometry and perform bounded MuJoCo compilation and five-step checks.

Inspection and preview; no motion permit

SSH experiment entry

Use an existing OpenSSH host configuration to run source-bound experiments remotely and verify returned evidence and transfer receipts.

Existing alias, key agent, and known_hosts required

Read-only robot diagnostics

Mock diagnostics and Universal Robots model, mode, and safety-status queries; this surface sends no motion commands.

Read-only; physical motion adapter pending

  1. Use the supported interactive installer and start the local workbench.
  2. Create or import a scenario and compare geometry and consequence results for four candidates.
  3. Run the experiment and inspect execution completion, evidence integrity, and scientific acceptance separately.
  4. Replay recorded motion and inspect events and the three runtime cursors.
  5. Request stop, drain, and approved recovery through the lifecycle, retaining the actually accepted tail.
  6. Export a signed ZIP and verify it with the independent verifier.

A recorded workbench run shows completed execution, evidence PASS, incomplete scientific acceptance, and slip together. The macOS App is a locally built research shell managed by a Python environment; distribution signing and notarization remain open.

Workbench validation and product records

Earlier study: observe the native policy.

The official LIBERO/Panda policy completes 58/100 in its original configuration. A later fresh-state confirmation uses MuJoCo 3.3.7 and reobserves every ten actions, completing 93/100 across ten tasks while retaining seven timeouts; the top-drawer task remains weakest at 6/10. The runs use different grids/configurations, and both record zero interventions.

93/100

Completed / 100 simulated episodes

Official LIBERO/Panda policy · MuJoCo 3.3.7 · predict 50 / execute 10 · Fresh states 30–39 for each task · Zero observer interventions.

95% Wilson interval: 86.25–96.57%. All seven timeouts remain in the results; the top-drawer task is still 6/10.

Ten tasks, with every denominator retained.
  1. Task 0010/10
  2. Task 0110/10
  3. Task 0210/10
  4. Task 039/10
  5. Task 046/10
  6. Task 059/10
  7. Task 0610/10
  8. Task 0710/10
  9. Task 089/10
  10. Task 0910/10

Official LIBERO/Panda policy · MuJoCo 3.3.7 · predict 50 / execute 10 · Fresh states 30–39 for each task · Zero observer interventions.

A separate grid/backend/horizon from 58/100; zero interventions. It is not a 35-point Sentinel gain, hardware performance, or SO100 task success.

Inspect the original research record
Fresh-state full-suite confirmation
Original Sentinel research figure: full suite confirmation
Original public research record · 02 October 2026

A separate grid/backend/horizon from 58/100; zero interventions. It is not a 35-point Sentinel gain, hardware performance, or SO100 task success.

Complete original-configuration record
Original Sentinel research figure: native policy and horizon
Original public research record · 02 October 2026

All 42 failures are retained; observe-only, distinct from the SO100 overlay, UR5e, and Sentinel intervention efficacy.

Inspect the original research record

Break a failure into testable questions.

Task 5 fails 10/10 originally. More frequent feedback yields 1/10; changing the complete MuJoCo version on fresh paired states yields 0/10 versus 7/10, with the control at 10/10 in both. A no-policy reset audit also finds an initial bowl-pose difference, without isolating one physical cause.

Original Sentinel research figure: backend version success
Complete MuJoCo version comparison · Same official model and predict 50 / execute 10 · Task 5: 0/10 versus 7/10; control task 0: 10/10 in both.
Open the original figure

Success and failure in the original configuration

Successful rollout · Task 0 / state 0 of the original 58/100 configuration · 83 actual actions · 4.15 s; not part of the fresh-state 93/100 cohort.
Failure replay · Task 5 / state 0 of the original 58/100 configuration · 280 actual actions · 14.05 s; a post-hoc exact replay, not an additional experiment.
Horizon, reset and complete-version comparisons

The horizon ablation uses 20 matched states and 40 rollouts; the reset audit compares 40 resets without loading a policy; the complete-version comparison uses a separate 20 matched states and 40 rollouts. They answer different questions and should not be combined into one improvement percentage.

Task 5 still fails 9/10; small-sample intervals include zero. This is not an overall model gain or an asynchronous chunk method.

Inspect the original research record
Original Sentinel research figure: reset version comparison
Original public research record · 02 October 2026

Shows a version difference before policy action; it neither proves reset-pose causation nor ranks physical accuracy.

Inspect the original research record

A complete-version benchmark-compatibility effect, without isolating one contact/reset cause. Task 5 still fails 3/10 under the compatibility version.

Inspect the original research record

Retain task utility alongside unsafe-action rejection.

The new 600-root router completes 300, rejects 300, and observes no unsafe selection, while fixed 4.8 s completes all 600. Across 27 broader stress cells, 4.8 s completes 324/324; 1.6 s completes 184/324 with 78 drops. The floor rule matches fixed slow for every root, at three times the action duration.

Original Sentinel research figure: routing utility
600 fresh simulated cases · Routing completes 300 and rejects 300; fixed 4.8 s completes 600. Rejection counts as incomplete.
Open the original figure
Low-friction replay · First prospectively ordered low-friction case in the new study · Stored simulator states · 10.2 s; the fast action drops, the slow action completes.
27 physical-condition cells and unobserved future friction
Original Sentinel research figure: physical stress
27 stratified condition cells and 324 cases · Fast actions complete 184/324 with 78 drops; fixed slow and the floor rule both complete 324/324, identically on every case.
Open the original figure

The floor rule equals fixed 4.8 s root by root, with no adaptive-duration gain; mu_min=.015 is declared. Finite strata do not qualify a continuous domain or hardware. The 3.2 s analysis is post-hoc.

Inspect the original research record

Across 100 matched cases, observed history, state and action bytes stay identical while future friction changes. Unsafe labels differ in 100/100 pairs for the 1.6 s action and 0/100 for the 4.8 s action.

Shows ambiguity from an unobserved future contact change; it does not prove static friction can never be identified through informative probing.

Inspect the original research record

Declared-profile routing, not automatic pixel OOD or friction estimation. All 300 unsupported mass/displacement rejections count as incomplete; fixed slow has stronger total utility.

Measure rechecking together with its cost.

Full and incremental decisions agree on all 60 fresh obstacle-field roots; parent-only checking falsely allows 20/23 unsafe roots. Fair rotated timing finds a 0.996 ms mean marginal saving with worse P95, supporting no general robot-speed claim.

Original Sentinel research figure: verification cost
60 fresh obstacle-field cases · Three rotated timing repeats with warmup excluded · Full and incremental decisions agree 60/60; mean marginal times are 85.351 ms and 84.356 ms.
Open the original figure

Mean marginal saving: 0.996 ms, paired difference interval −1.728 to −0.291 ms. Total-cost P95 rises from 223.214 ms to 226.218 ms, worse for the incremental path. Only three cases concern collisions; other hazards mainly use designed joint and tracking conditions.

Inspect the original research record
Mechanism film · Rerendered stored trajectories from the earlier 180-case study · 19.95 s. Red rejected proposals are counterfactual and were not dispatched.

Earlier studies: mechanisms, models and scenes.

These earlier studies retain their own samples and configurations. Offline SO100 action reconstruction, the 180-case UR5e comparison and WorldGuard scene studies are reported separately from the new native Panda tasks and fresh cost experiment.

Experimental setup, full results and original figure

The UR5e study compares 180 constructed, same-root simulation cases. Independent higher-resolution review labels 139 unsafe and 41 safe; each method faces the same plan changes.

Parent-only checking falsely allows 125 unsafe cases; full final validation and the incremental-with-fallback path allow none in this set. These are results for the constructed simulations, rather than a guarantee across robot settings.

Decisions and the cost of rechecking

False allows / 139 unsafe cases

Parent only
125/139
Full final validation
0/139
Incremental + fallback
0/139

Mean measured total cost

Parent only
101.15 ms
Full final validation
48.74 ms
Incremental + fallback
147.65 ms

The incremental path takes 147.65 ms including the parent check, versus 48.74 ms for full final validation. This result does not establish a speedup.

Experimental setup, full results and original figure

Six scenario classes with 30 roots each produce 180 same-root comparisons. A separately implemented reviewer uses 0.005 rad static sampling and 1 ms MuJoCo steps, labeling 139 roots unsafe and 41 safe.

UR5e same-root comparison
PolicyAllowed / 180False allows / 139 unsafeFalse rejects / 41 safeMean observed wall time
Parent only166/180125/1390/41101.15 ms
Full final validation41/1800/1390/4148.74 ms
Incremental + fallback41/1800/1390/41147.65 ms
Reject transforms30/1800/13911/41No validation call

106 roots reuse the static prefix and 74 fall back fully. The incremental stage averages 46.50 ms plus 101.15 ms for parent validation: 147.65 ms total, slower than full validation's 48.74 ms. Dynamic replay always starts from frame zero.

Seven representative cases, 720p, 33.75 s including titles and final-state holds. The 287 stored joint-state frames run at 1× for 14.35 s. Rendering reruns neither physics nor the gate.

False allows, false rejects, and parent-inclusive validation cost for 180 constructed UR5e roots.
Parent-only falsely allows 125/139; full and incremental both 0/139. Incremental total cost is 147.65 ms versus 48.74 ms full, without a speedup claim.
Original UR5e review results
Data splits, model package and evaluation settings

The SmolVLA study uses official SO100 PickPlace recordings to fine-tune the action expert and related projections, preserving separate training, selection and held-out evaluation splits. The model package, loader and evaluation records are published on Hugging Face.

63.46%

Lower normalized action MAE versus the frozen base model

Five held-out episodes · Recorded-action reconstruction

The metric concerns reconstruction of recorded actions. Episodes are the independent units; overlapping windows are not independent successful tasks.

Data splits, model package and evaluation settings

Fine-tune the action expert and state/action/time projections over a pinned SmolVLA base and frozen SmolVLM2 backbone. Training runs for 5,000 updates, development selects update 3,750, and 99,880,992 parameters are trainable.

Official SO100 PickPlace: 50 episodes, 19,631 frames, 30 Hz, six joint/gripper fields, and top/wrist cameras. The study uses 30 train / 5 dev / 10 cal / 5 test episodes, with normalization fitted on training data only.

One paired held-out evaluation with identical sampling noise lowers normalized MAE by 63.46%. Five held-out episodes produce 1,926 overlapping windows, each predicting a 50×6 action chunk; the independent unit remains the episode.

SmolVLA held-out action reconstruction
PolicyNormalized action MAENative-field action MAE
Pinned base0.67316613.321098
Dev-selected fine-tune0.2459844.47285

The Hub publishes a 21-file overlay, loader, model card, and evaluation record. The pinned original base and backbone are required; one dev-only [1,50,6] example passed a bitwise equality check.

This result measures reconstruction of recorded actions, not robot task success. The data does not declare common physical units; SO100 overlay results cannot be attributed to the separate UR5e simulator study.

RTX 4090 D (24 GB), four threads, batch one: in-memory images through unnormalized actions measure P50 229.26 ms / P95 236.90 ms over 20 synchronized calls. Video reading, transport, and actuation are excluded.

Left: SmolVLA action MAE falls 63.46% over five held-out episodes. Right: matched calibration reduces unsafe selections from 47 to 0 on the same 500-root camera stress set.
Two distinct experiments: recorded-action reconstruction on the left, simulator camera-profile recalibration on the right. The right-side predictions are unchanged and actions shift from 1.2 to 1.6 s.
Six scenarios and their supported scope

WorldGuard tests consequence prediction under distribution shift. Cameras, payload, friction and the supported action range change the conditions for a decision. The study retains failures alongside selection coverage and simpler comparators.

WorldGuard unsafe selections and XY coverage across six fresh scenarios, retaining low-friction and camera failures.
100 fresh roots per scenario. Under camera shift, the image gate selects 7/100 unsafe actions; every original low-friction candidate is unsafe. The 1.35× move is an out-of-contract bare-model diagnostic; the product path rejects all.

A changed camera needs matched calibration.

On the earlier fixed 500-root stress set, shifting cameras with the original calibration produced 47/500 unsafe selections. Matching the dev/cal profile reduced this to 0/500 by selecting the slower 1.6 s action for all roots. Predictions are unchanged, and XY coverage is 94.0%, below the nominal 95%.

Six scenarios and their supported scope

W0/W1/W2 train 36 members across numerical dynamics, contact consequences, visual objects, and real SO100 joint recordings. Stronger simple comparators remain visible: physical identification beats the strict learned numerical model, while adding images increases normalized SO100 joint-prediction error by 10.0%.

Six scenarios, each with 100 fresh roots and four sibling actions per root. Counts below are unsafe selections / 100.
ScenarioFast geometryFixed 1.6 sState WorldGuardImage WorldGuard
Nominal59000
Hidden low friction100100100100
Low mass76000
High mass75000
Camera shift62007
1.35× displacement (bare-model diagnostic)89100

The 1.35× displacement exceeds the action contract. The declared product path returns MODEL_UNKNOWN and rejects all 100 roots; bare-model values describe degradation, not supported-profile authorization.

Camera and preprocessing identities belong with the model, PCA, action family, and calibration. Under hidden low friction, every original candidate is unsafe; prediction or recalibration cannot create a safe action.

WorldGuard scenario report

A safe action must actually exist.

A separate low-friction profile studies 100 fresh simulated cases, extending a 1.6-second action family with slower candidates. The 4.8-second plan completes 100/100 tray tasks with no drops, at three times the action duration.

Outcomes and target-acceleration screening for 1.6, 3.2, and 4.8 s plans in a new low-friction profile.
100 fresh roots, separate 5 s profile, configured μ_min=0.015. The 4.8 s plan completes 100/100 tray tasks with zero drops at 3× duration; 3.2 s is also safe here but the rule rejects it.
Full low-friction action-family results

A separate 5 s physical profile tests 1.6, 3.2, and 4.8 s plans over 100 fresh low-friction roots. The fixed 1.6 s plan has 100/100 unsafe outcomes, 58/100 drops, and 0/100 tray completions; 4.8 s has 0/100 unsafe outcomes, 0/100 drops, and 100/100 tray completions, at three times the action duration.

Screening uses configured μ_min=0.015, without measuring friction. The 3.2 s plan is also safe in this sample but the conservative rule rejects it. Zero/100 still has a 3.70% Wilson upper bound; this is a new, undeployed simulator profile.

Low-friction action-family report

Read, try and reproduce.

Start with the project site, then follow the code, models and experiment records for closer inspection.

Local installation and startup

The interactive installer offers full and core profiles, creates a managed Python environment and can build the macOS App. Python 3.10+ is required; the service binds only to 127.0.0.1. Windows provides a CLI installation entry. Native policy inference uses a separately prepared LeRobot / LIBERO environment.

python -m pip install -e ".[physics]"
python -m sentinel_evc serve --data-dir runs/workbench --port 8765
What this release validates

Each study retains commands, source/environment identities, input revisions, model/calibration digests, and per-root or per-episode artifacts. Signatures establish integrity relative to the chosen public key; replay shows stored data without recomputing the experiment.

  • Numeric core, fixed contact, GPU models, UR5e, and LIBERO protocols are separately recorded profiles.
  • Robot entry is read-only; imported-model inspection grants no motion authority.
  • Zero observed failures, prediction coverage, task completion, and physical incidents are distinct metrics.
  • Signatures establish neither sensor honesty, physical execution causation, nor safety certification.
  • ROS 2 motion, physical actuation, mTLS, production PKI, and notarized distribution remain open.

Checked 03 October 2026 PDT · Earlier studies retain their frozen revisions

GitHub 009a24c · Hugging Face 1fe97cc

Each result retains its own configuration, sample and scope. Code: MIT. Model and data: Apache-2.0. Official UR5e assets retain their original licenses.

Formal research release · 02 October 2026

KEEP IN TOUCH

Continue the conversation about that final step.

Let’s discuss action verification, research replication and future collaboration.

Contact & conversations