Physical AI Vla Trajectory Fork 029 - Raw Research Artifact
Generated at: 2026-07-17T17:42:28Z Classification: Research-only Status: fail
Causal Result
Section titled “Causal Result”Paired counterfactual forks: 14 across 2 tasks and 3 trajectory phases.
| Metric | Result |
|---|---|
| Historical-to-VLA action MAE | 0.1415 |
| Post-action normalized state drift p50 | 0.0113 |
| Post-action front-pixel MAE p50 | 0.00413 |
| Branch minus control confidence | -0.0002 |
| Control eligibility | 94.6% |
| Historical branch eligibility | 94.6% |
| Eligibility drop | 0.0% |
| Action-error/state-drift correlation | 0.960 |
By Region
Section titled “By Region”| Region | Forks | Action MAE | State drift | Pixel MAE | Confidence delta | Control eligible | Branch eligible |
|---|---|---|---|---|---|---|---|
| broad | 5 | 0.1500 | 0.0135 | 0.00488 | -0.0001 | 100.0% | 100.0% |
| middle | 5 | 0.1500 | 0.0135 | 0.00488 | -0.0001 | 100.0% | 100.0% |
| selected | 4 | 0.1201 | 0.0101 | 0.00484 | -0.0004 | 81.2% | 81.2% |
Acceptance
Section titled “Acceptance”| Check | Status |
|---|---|
| coverage | pass |
| dose response | pass |
| eligibility effect | fail |
| restore exact | pass |
| search p95 | pass |
| trajectory shift | pass |
| aws resources deleted | pass |
ZeptoDB search p50/p95: 5.139/5.486 ms.
Interpretation
Section titled “Interpretation”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.