Skip to content

Physical AI Vla Closed Loop 027 - Raw Research Artifact

Generated at: 2026-07-17T08:57:14Z Classification: Research-only Suite: libero_10 VLA: HuggingFaceVLA/smolvla_libero at 6721902bc4d61e50a3bfdb11dfb4cb626f05d102 Retrieval encoder: google/siglip-base-patch16-224

Direct VLA and the ZeptoDB-routed policy ran from paired LIBERO initial states. Success is reported by the real simulator. This bounded run uses one initial state per task and is not a production safety claim.

TasksInit states/taskMax stepsMemoriesThresholdGuard
1015201900.8908412 high-confidence, max 4 skips
PathSuccessesSuccess rateSteps
Direct VLA5/1050.0%4291
ZeptoDB routed5/1050.0%4286

Paired regressions: 0; paired improvements: 0.

PathVLA callsSkipsMean decision msp95 msGPU total msEnergy J
Direct VLA42910454.169469.2411877964.838200326.4
ZeptoDB routed42860469.756482.1331927346.806204915.6
  • VLA call reduction: 0.1%.
  • Fallback rate: 100.0%.
  • Mean decision-latency reduction: -3.4%.
  • Online GPU-time reduction: -2.6%.
  • ZeptoDB search p50/p95: 2.425/2.671 ms.
TaskDirect successRouted successDirect stepsRouted stepsRouted VLASkips
0yesyes4514514510
1yesyes2732732730
2yesyes2502452450
3yesyes5075075070
4nono5205205200
5yesyes2102102100
6nono5205205200
7nono5205205200
8nono5205205200
9nono5205205200
CriterionStatus
Direct VLA succeeds on at least one taskpass
Fallback rate is between 5% and 95%fail
Mean decision latency reduced by at least 15%fail
At most one paired task regressionpass
Real LIBERO closed-loop steps executedpass
Routed success rate no more than 10pp below direct VLApass
VLA calls reduced by at least 30%fail
ZeptoDB exact-search p95 below 30 mspass
Temporary AWS resources deletedpass

Overall status: fail.

This run tests real closed-loop behavior for a bounded deterministic slice. It does not establish confidence intervals, drift robustness, collision safety, or readiness to control a physical robot.