Skip to content

Physical AI Vla Skip Region 028 - Raw Research Artifact

Generated at: 2026-07-17T16:04:16Z Classification: Research-only Status: fail Suite: libero_10 VLA: HuggingFaceVLA/smolvla_libero at 6721902bc4d61e50a3bfdb11dfb4cb626f05d102 Retrieval encoder: google/siglip-base-patch16-224

The staged run stopped at closed_loop_pilot_gate_failed.

Shadow steps: 801; successful pilot baselines: 1/2. ZeptoDB search p95: 2.593 ms.

ConfidenceMin marginSkip rateMean action MAEp95 action MAEProjected latency reduction
0.7600.01028.8%0.16400.261225.8%

The broadest shadow-safe candidate was confidence 0.68 with no margin minimum. It projected 63.9% skips and 60.9% lower latency with mean/p95 action MAE of 0.1925/0.3558, but it was outside the planned 20% to 30% calibration target and was not executed in closed loop.

  • Paired regressions: 0.
  • Actual skip rate: 13.3%.
  • Actual latency reduction: not comparable; this run’s routed timer included the simulator step while the shadow baseline did not.

No full ten-task routed run was started after the failed gate.

CriterionStatus
Task-partitioned retrievalpass
Viable shadow skip regionpass
No paired pilot regressionpass
Actual pilot skip rate at least 20%fail, 13.3%
Comparable actual latency reduction at least 15%not evaluated
ZeptoDB exact-search p95 below 30 mspass, 2.593 ms
Temporary AWS resources deletedpass

Experiment 028 found an offline-on-policy shadow region, but not a stable closed-loop skip region. Executing retrieved actions changed subsequent observations enough that eligibility fell from 28.8% to 13.3%. This is useful potential evidence, not a demonstrated compute-saving policy.

The pilot decision-timer scope mismatch was found after cleanup. Skip counts, action-error calibration, simulator success, regression count, and ZeptoDB latency remain valid; the reported pilot latency reduction must not be used. The harness now stops its decision timer before env.step() for a future run.

Temporary c7i.xlarge and g6e.xlarge instances reached terminated; the experiment namespace, NodePools, and NodeClaims were absent; shared bench NodePool CPU limits were zero.