Skip to content

Physical AI Vla Trajectory Fork 029 - Raw Research Artifact

Generated at: 2026-07-17T17:42:28Z Classification: Research-only Status: fail

Paired counterfactual forks: 14 across 2 tasks and 3 trajectory phases.

MetricResult
Historical-to-VLA action MAE0.1415
Post-action normalized state drift p500.0113
Post-action front-pixel MAE p500.00413
Branch minus control confidence-0.0002
Control eligibility94.6%
Historical branch eligibility94.6%
Eligibility drop0.0%
Action-error/state-drift correlation0.960
RegionForksAction MAEState driftPixel MAEConfidence deltaControl eligibleBranch eligible
broad50.15000.01350.00488-0.0001100.0%100.0%
middle50.15000.01350.00488-0.0001100.0%100.0%
selected40.12010.01010.00484-0.000481.2%81.2%
CheckStatus
coveragepass
dose responsepass
eligibility effectfail
restore exactpass
search p95pass
trajectory shiftpass
aws resources deletedpass

ZeptoDB search p50/p95: 5.139/5.486 ms.

The fork begins from an identical restored MuJoCo state. One branch uses the direct VLA action and the other uses one retrieved historical action; both then use deterministically paired VLA actions for up to four steps. This isolates short-horizon trajectory shift but is not physical-robot or multi-seed production evidence.

The tested cause is not supported. A single historical action did move the trajectory: normalized state drift p50 was 0.0113, front-pixel MAE p50 was 0.00413, and action error strongly correlated with state drift (r=0.960). However, the shift did not reduce retrieval eligibility. Control and historical branches were both eligible on 94.6% of post-action observations, and mean branch-minus-control confidence was only -0.0002.

The Experiment 028 skip-rate collapse therefore cannot be attributed to the immediate four-step confidence effect of one historical action in these forks. Remaining candidates include two consecutive historical actions, sampling bias toward stable first-eligible pockets, and stochastic SmolVLA RNG stream misalignment: the routed path did not consume a policy sample on a skip, so its next fallback did not use the baseline step’s random stream.

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