Toy closed-loop utility study v0.1

Question

Does a future scorer with better offline candidate ranking also obtain better return when the selected action is actually executed and the scorer must replan from the resulting state?

This study is an evaluator validation on exact analytic dynamics. It is not evidence that the released StarWAM checkpoint improves closed-loop LIBERO performance: that adapter does not expose action-conditioned candidate futures, so WAMProbe does not fabricate them.

Fixed protocol

  • benchmarks: BlockPush-2D v0.1 and Gripper-Catch v0.1;
  • shared contexts: 12 per benchmark, suite seed 7;
  • legal candidate set: the benchmark's full fixed candidate set at every cycle;
  • prediction horizon: 6 steps for BlockPush, 5 for Gripper-Catch;
  • episode budget: one benchmark horizon;
  • execute prefix: one true-dynamics step before observing and replanning;
  • score horizon: prediction prefix capped by remaining episode steps;
  • ties: stable benchmark candidate order;
  • uncertainty: 1,000 context-block bootstrap resamples;
  • success: match the greedy simulator-future scorer's final return within 1e-9.

Controllers include oracle, seeded noisy, copy-last, wrong-direction and action-agnostic future scorers, plus a simulator-future scorer, a fixed action-only policy and deterministic random selection. Offline CRC and Top-1 Regret are intentionally null for controllers that do not emit candidate futures.

Reproduce

wamprobe closed-loop-study \
  --benchmark all \
  --contexts 12 \
  --seed 7 \
  --execute-prefix 1 \
  --resamples 1000 \
  --output runs/closed-loop-study-v0.1

The canonical JSON SHA256 is 19d2a0108fc19580da4e81f15226a4706a40c9d2c0513857d6bc8dc9000189b3.

Results

Benchmark Future scorer Offline CRC Offline regret Closed return Success
BlockPush oracle 1.0000 0.0000 0.0000 1.0000
BlockPush noisy 1.0000 0.0000 0.0000 1.0000
BlockPush copy-last 0.0000 0.3000 -0.3000 0.0000
BlockPush wrong-direction -0.2500 0.3000 -0.3000 0.0000
BlockPush action-agnostic 0.0000 0.3000 -0.3000 0.0000
Gripper-Catch oracle 1.0000 0.0000 1.0000 1.0000
Gripper-Catch noisy 1.0000 0.0000 0.9967 0.9167
Gripper-Catch copy-last 0.0000 1.0000 0.0000 0.0000
Gripper-Catch wrong-direction 0.0000 1.0000 0.0000 0.0000
Gripper-Catch action-agnostic 0.0000 1.0000 0.0000 0.0000

Across the five future-scorer profiles, offline CRC versus mean closed-loop return has Pearson correlation 0.9855 for BlockPush and 1.0000 for Gripper-Catch. Offline Top-1 Regret versus return is -1.0000 for both. No comparable future-scorer pair reverses the offline-CRC/return ordering.

Interpretation and limits

The result confirms that the evaluator connects offline action ranking to actual repeated execution on two controlled tasks. It does not establish a population-level correlation: there are only five deliberately separated reference profiles, three weak profiles tie on return, and the contexts come from analytic generators rather than independent robot task distributions. The action-only fixed policy also reaches the task optimum, showing that a world model is not necessary on these easy generated tasks.

The first implementation used the full prediction endpoint even when fewer episode steps remained. That made an exact model overrun the goal near termination. The released protocol caps the scoring prefix by the remaining budget, and regression tests preserve this MPC semantics. Larger LIBERO studies must likewise declare prediction horizon, execution prefix, reward horizon, terminal handling and candidate generator before comparing controllers.