docs/public_bridge_smoke_run_log_2026-04-01.md

Public Bridge Smoke Run Log

Date: 2026-04-01 UTC

Completed public proxy evidence

• Occlusion proxy already completed earlier on PickClutterYCB-v1.

• Best current occlusion report:

  • /workspace/workspace/reports/maniskill_pickclutter_smoke_v5_eval_tuned_softerpref/public_benchmark_package_summary.json
  • trunk_only_ft=0.04
  • adapter_noop=0.04
  • adapter_active_ft=0.62
  • delta_active_vs_trunk=+0.58
  • 95% CI [0.44, 0.72]
  • intervention_rate=1.0
  • non_base_selection_rate=1.0

• Bag proxy completed on the public ManiSkill bridge basket scene proxy.

• Bag report directory:

  • /workspace/workspace/reports/maniskill_bag_bridge_smoke_v1

• Bag result summary:

  • trunk_only_ft=0.32
  • adapter_noop=0.00
  • adapter_active_ft=0.48
  • delta_active_vs_trunk=+0.16
  • delta_active_vs_trunk_ci95=[-0.04, 0.34]
  • intervention_rate=1.0
  • non_base_selection_rate=1.0
  • bag track signs_of_life=true
  • package-level headline gate remains false at this single-seed smoke scale

Cloth proxy definition

• Public scene proxy:
  • PutSpoonOnTableClothInScene-v1
• Fixed hidden-state initialization:
  • spoon pose [-0.235, -0.094, 0.8748]
  • cloth pose [-0.235, -0.075, 0.885]
• Deterministic valid-seed filter:
  • accept only seeds whose initialized hidden state is below the visibility gate and solvable by scripted reveal+retrieve
• Reveal macros corrected to push-style actions:
  • lift_edge = front push in +y
  • separate_layer = side push in +x
• Cloth success metric corrected:
  • based on spoon displacement from its own hidden start plus visibility
  • no longer credits success merely because the cloth flies away

Important runner fixes already landed

• File:
  • /workspace/workspace/VLAarchtests3_export/code/VLAarchtests2_code/VLAarchtests/code/reveal_vla_bimanual/eval/run_maniskill_bridge_retrieval_smoke.py
• Fixed:
  • cloth hidden-state initialization
  • cloth seed filtering and split reuse via episode_splits.json
  • post_bundle missing in cloth collect success check
  • bridge smoke loss weights aligned to current LossWeights
  • adapter trainable parameter prefixes aligned to working pickclutter runner
  • zero-depth layout changed to channel-first
  • cached dataset normalizer added for old channel-last depth tensors

Live status when this note was written

• Bag process is complete.
• Cloth process is still collecting the train split in the original long-running session.
• The long-running cloth process was started before the later loss-weight and depth-layout fixes, so it is expected to finish collection and then crash at training start.
• After it writes train.pt and val.pt, restart cloth with:

python /workspace/workspace/VLAarchtests3_export/code/VLAarchtests2_code/VLAarchtests/code/reveal_vla_bimanual/eval/run_maniskill_bridge_retrieval_smoke.py --task cloth --skip-collection

• If trunk checkpoint already exists by that point and only adapter needs rerun:

python /workspace/workspace/VLAarchtests3_export/code/VLAarchtests2_code/VLAarchtests/code/reveal_vla_bimanual/eval/run_maniskill_bridge_retrieval_smoke.py --task cloth --skip-collection --reuse-checkpoints

Cloth restart correction

• The corrected cloth restart reached adapter training and failed in rollout supervision because the cached cloth public proxy authored 7 candidate targets while the decoder always allocates 8 proposal slots.
• Fix landed in:
  • /workspace/workspace/VLAarchtests3_export/code/VLAarchtests2_code/VLAarchtests/code/reveal_vla_bimanual/eval/run_maniskill_bridge_retrieval_smoke.py
• Correction:
  • cached bridge samples now normalize channel-last depth tensors as before
  • cached candidate-aligned tensors now also pad from 7 -> 8 slots before loading
  • padding cycles the non-base candidates first, which preserves the collected cloth episodes and avoids recollection
• Verified locally before restart:
  • normalized cloth candidate_action_chunks is (8, 8, 14)
  • normalized cloth candidate_rollout_support_mode is (8, 5)
  • one real adapter_active_ft training step and one real validation loss pass both completed without the previous shape error

Cloth result

• Report directory:
  • /workspace/workspace/reports/maniskill_cloth_bridge_smoke_v1
• Final cloth smoke summary:
  • trunk_only_ft = 0.04
  • adapter_noop = 0.04
  • adapter_active_ft = 0.10
  • delta_active_vs_trunk = +0.06
  • delta_active_vs_trunk_ci95 = [-0.04, 0.16]
  • intervention_rate = 0.3369
  • non_base_selection_rate = 0.2674
• Interpretation:
  • cloth proxy is positive and adapter-specific in this single-seed smoke because adapter_noop stayed flat while adapter_active_ft improved
  • effect size is modest and not yet statistically clean in this smoke protocol

Combined three-track proxy suite

• Combined report:
  • /workspace/workspace/reports/public_proxy_suite_smoke_v1/combined_summary.json
  • /workspace/workspace/reports/public_proxy_suite_smoke_v1/combined_summary.md
• Current three-track smoke evidence:
  • occlusion proxy positive and adapter-specific
  • bag proxy positive and adapter-specific
  • cloth proxy positive and adapter-specific