Experiment 029: Physical AI VLA Trajectory Fork On EKS

Date: 2026-07-17 Status: Research complete - tested cause not supported Classification: Research-only

Goal

Determine whether the Experiment 028 closed-loop skip-rate collapse is caused by a retrieved historical action moving the robot into observations with lower retrieval confidence and eligibility.

Hypothesis

Starting from an identical MuJoCo state, one historical action will create measurable robot-state and front-camera divergence from the direct VLA control branch. Over the following four paired VLA steps, the historical branch will show at least a ten-percentage-point eligibility drop or a mean confidence drop of at least 0.01.

Scope

• Pinned HuggingFaceVLA/smolvla_libero and SigLIP models from Experiments 026-028.
• The same 190 task-partitioned historical action memories as Experiment 028.
• LIBERO-10 tasks 0 and 5, seed base 28, and at most 520 shadow steps.
• Three Experiment 028 candidate regions:
  • broad: confidence 0.68, minimum margin 0.
  • middle: confidence 0.72, minimum margin 0.005.
  • selected: confidence 0.76, minimum margin 0.01.
• At most one first eligible snapshot per task, phase, and region.
• One historical-action intervention per fork, followed by at most four deterministically paired VLA recovery steps.

This experiment isolates short-horizon simulator causality. It does not test a production router, broad task success, physical safety, or multi-seed robustness.

Procedure

1. Refresh the balanced LIBERO manifest before creating AWS resources.
2. Start temporary c7i.xlarge and g6e.xlarge EKS Auto Mode NodePools.
3. Load pinned models and insert memories into task-specific ZeptoDB namespaces.
4. Run direct VLA shadow trajectories on tasks 0 and 5.
5. When a region is eligible for two consecutive observations, retain the first MuJoCo state snapshot in each early, middle, or late phase.
6. Create two fresh LIBERO environments and restore both to the same snapshot. Verify normalized state error at most 1e-8 and exact front pixels.
7. Apply the recorded direct VLA action to the control environment and the retrieved historical action to the branch environment.
8. At offsets 1 through 4, measure normalized robot-state MAE, front-pixel MAE, ZeptoDB/SigLIP confidence, and region eligibility.
9. Use the same policy seed for paired VLA recovery actions at each offset.
10. Test whether intervention action error correlates with subsequent state drift and whether the historical branch loses confidence or eligibility relative to control.
11. Recover compact results from /dev/termination-log and delete all temporary resources.

Cause-Support Criteria

Criterion	Required
Fork coverage	at least 6, 2 tasks, 2 phases
Paired restored state	state MAE <= 1e-8 and pixel MAE = 0
Post-action trajectory shift	normalized state MAE p50 >= 0.01
Eligibility effect	eligibility drop >= 10pp or confidence drop >= 0.01
Dose-response support	action-error/state-drift correlation >= 0.30
ZeptoDB search p95	below 30 ms
Temporary AWS resources deleted	pass

Trajectory shift plus an eligibility effect is the causal core. Dose response is reported separately because the bounded fork count may not support a stable correlation estimate.

Failure Behavior

Dependency, model, snapshot restore, simulator, ZeptoDB, timeout, compact result, or cleanup errors stop the run. Fewer than six eligible fork targets stop before counterfactual execution. Cleanup runs before failure investigation, and pod logs are not used as the result channel.

Command

tests/k8s/run_physical_ai_vla_trajectory_fork.sh

Result

Fourteen counterfactual forks covered both tasks and all three trajectory phases. Restoration was exact: maximum paired normalized state and front-pixel MAE were both zero. One historical action produced normalized state drift p50 0.0113 and front-pixel MAE p50 0.00413. Historical-to-VLA action error strongly predicted state drift (r=0.960).

The required retrieval effect did not occur. Control and historical branches were each eligible on 94.6% of the 56 post-action observations, for a zero-percentage-point eligibility difference. Mean branch-minus-control confidence was -0.0002, far below the required 0.01 drop. Exact ZeptoDB search p95 was 5.486 ms.

The causal hypothesis therefore failed: one historical action does shift the short-horizon state and image trajectory, but that shift does not explain Experiment 028’s eligibility collapse. Temporary Kubernetes resources were absent after cleanup, both temporary EC2 instances terminated, and shared bench NodePool CPU limits returned to zero. See docs/research/results/physical_ai_vla_trajectory_fork_029.md.

Interpretation

The retrieval embedding is robust to the measured one-action trajectory shift. Experiment 028’s lower routed skip rate is more likely to involve repeated historical-action bursts, which this experiment intentionally excluded, or stochastic policy-stream divergence. In Experiment 028, a skip did not execute SmolVLA and therefore did not consume the random samples that the direct path used at that control step; a later fallback was no longer sampling from the same per-step random stream.

Next Research Step

Before changing thresholds, seed SmolVLA by absolute task/episode/control-step for every fallback so direct and routed paths remain paired after skips. Then compare one versus two consecutive historical actions and sample eligible points across each phase instead of only the first stable pocket.