VLAarchtests2

• current code under VLAarchtests/
• current third-party baseline code under third_party/
• current baseline runs, replay artifacts, demo roots, and released checkpoint material under baselines/
• current training outputs and checkpoints under outputs/
• current logs under reports/
• environment recreation files under environment/
• raw results and change/test logs at the repo root
• the previous repo README under history/VLAarchtests_previous_README.md
• the active handoff file under handoff/instructions4.md

Top-Level Contents

• VLAarchtests/
  • code, tests, configs, generated configs, reports, checkpoints, and proxy datasets from the current runpod workspace
• third_party/AnyBimanual/
  • local AnyBimanual checkout used for the official overlap baseline branch, including local compatibility patches
• baselines/
  • released AnyBimanual checkpoint material
  • overlap replay artifacts
    • HF export packaging note: baselines/AnyBimanual_overlap_replay/multi/ is sharded into subdirectories to satisfy the Hub 10000 files per directory limit
  • overlap run directories
  • local subset3 demo roots used by the overlap branch
• outputs/
  • RLBench training outputs and checkpoints used by the current anchor, RVT, dual-push, and elastic-controller branches
• reports/
  • training and evaluation logs copied from /workspace/reports
• environment/
  • machine snapshot, package lists, and setup helpers
• history/
  • copied previous-repo README
• handoff/
  • active sprint instruction file
• RESULTS_RAW.md
  • raw result tables and final official overlap eval outputs
• CHANGE_AND_TEST_LOG.md
  • file-level change log and executed test commands
• MODEL_AND_ARTIFACT_INDEX.md
  • staged directory map with main artifact roots

Previous Repo Coverage

• history/VLAarchtests_previous_README.md

• compact, spatial, compact-phase, and spatial-phase proxy branches
• earlier RLBench direct-policy and kNN runs
• environment recreation files
• prior raw result tables

Current Session Additions

• VLAarchtests/artifacts/reports/sprint_v7_summary/
• VLAarchtests/artifacts/reports/sprint_v7_followup/
• VLAarchtests/artifacts/reports/selector_finetune_v7_iterations/
• VLAarchtests/artifacts/reports/selector_finetune_v7_iter6/
• VLAarchtests/artifacts/reports/selector_finetune_v7_iter7/
• VLAarchtests/artifacts/reports/selector_finetune_v7_iter8/
• VLAarchtests/artifacts/reports/task_routed_proxy_v1/
• VLAarchtests/artifacts/reports/rlbench_general_debug_20260330/
• VLAarchtests/artifacts/reports/anybimanual_overlap_baseline_20260330/
• VLAarchtests/artifacts/reports/bag_mode_specialization_20260330/
• VLAarchtests/artifacts/reports/general_task_anchor_20260330_dual_push_buttons/
• VLAarchtests/artifacts/reports/rvt_overlap_branch_20260330/
• VLAarchtests/artifacts/reports/rvt_overlap_branch_fixedbounds_20260330/
• VLAarchtests/artifacts/reports/dual_push_nonzero_branch_20260330/
• VLAarchtests/artifacts/reports/dual_push_full_arch_hybrid_20260331/

Raw Results Snapshot

Proxy sprint v7

• VLAarchtests/artifacts/reports/sprint_v7_summary/reveal_sprint_summary_compact.json

• base model mean success: 0.28
• base per-task: foliage 0.39, bag 0.31, cloth 0.14
• random mean success: 0.43333333333333335
• candidate0 mean success: 0.2
• oracle mean success: 0.4066666666666667
• scripted mean success: 1.0

Eval-time ablations

• VLAarchtests/artifacts/reports/sprint_v7_summary/reveal_sprint_summary_compact.json

• no_planner: 0.2
• no_memory: 0.3233333333333333
• no_task_conditioning: 0.28
• no_geometry: 0.27
• no_camera_pose: 0.29333333333333333

Selector checkpoints

• VLAarchtests/artifacts/reports/selector_finetune_v7_iter6/default/reveal_benchmark.json
• VLAarchtests/artifacts/reports/selector_finetune_v7_iter7/full_fixed_default/reveal_benchmark.json
• VLAarchtests/artifacts/reports/selector_finetune_v7_iter8/bag_fixed_default/reveal_benchmark.json
• VLAarchtests/artifacts/reports/task_routed_proxy_v1/summary.md

• iter6 mean success: 0.4566666666666667
  • foliage 0.46, bag 0.4, cloth 0.51
• iter7 mean success: 0.4666666666666666
  • foliage 0.4, bag 0.41, cloth 0.59
• iter8 bag-only fixed slice: 0.41
• routed controller mean success: 0.48666666666666664
  • routing rule: foliage -> iter6, bag -> iter8, cloth -> iter8
  • per-task: foliage 0.46, bag 0.41, cloth 0.59

Real baseline compare on proxy suite

• VLAarchtests/artifacts/reports/real_baseline_compare_v7_full/reveal_benchmark.json

• baseline_rgbd_stage3 mean success: 0.31
  • foliage 0.21, bag 0.15, cloth 0.57
• iter5_selector mean success: 0.45
  • foliage 0.44, bag 0.4, cloth 0.51

RLBench recovered push-box comparator

• reports/rlbench_general_debug/rlbench_push_box_fair_step1_final_knn_ep10_x99_res224_len180_train80_fixed/bimanual_push_box/rollout_eval.json
• reports/rlbench_general_debug/rlbench_push_box_historical_step1_knn_ep10_x99_res224_len180_train80_fixed/bimanual_push_box/rollout_eval.json

• current fair-step1 final mean success: 0.7
• current fair-step1 final successes:
  • [1.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 1.0, 1.0]
• historical push-box control mean success: 0.4
• historical push-box control successes:
  • [0.0, 0.0, 0.0, 1.0, 1.0, 0.0, 0.0, 1.0, 1.0, 0.0]

Official AnyBimanual overlap branch

• baselines/AnyBimanual_overlap_runs/peract_bc_subset3_overlap_smoke200_fixpretrain_nowandb3/PERACT_BC/seed0/training.log
• reports/anybimanual_subset3_overlap_resume1000_eval.log

• global step 300: loss 40.91718
• global step 400: loss 33.26684
• global step 500: loss 36.07054
• global step 600: loss 35.32345
• global step 700: loss 28.50959
• global step 800: loss 23.60169
• global step 900: loss 15.28901
• run reached weights/1000 and the train exited cleanly

• source log: reports/anybimanual_subset3_overlap_resume1000_eval.log
• summary files:
  • VLAarchtests/artifacts/reports/anybimanual_overlap_baseline_20260330/resume1000_summary/summary.md
  • VLAarchtests/artifacts/reports/anybimanual_overlap_baseline_20260330/resume1000_summary/summary.json
• local last complete step: 1000
• local mean success: 0.16
• local per-task success:
  • coordinated_push_box: 0.0
  • coordinated_lift_ball: 0.0
  • dual_push_buttons: 0.48
• local per-task return:
  • coordinated_push_box: 0.0
  • coordinated_lift_ball: 0.0
  • dual_push_buttons: 12.0
• public best overlap step in the local summary: 60000
• public best mean success in the local summary: 0.6933333333333334

Validated general-task anchor: dual_push_buttons

• VLAarchtests/artifacts/reports/general_task_anchor_20260330_dual_push_buttons/summary.json
• baselines/AnyBimanual_release_eval_anchor/perlf_release_dual_push_buttons_ep25/PERACT_BC/seed0/eval_data.csv

• public AnyBimanual release, step 60000: success 0.96, return 24.0, length 21.56
• local official single-task eval, step 60000, 25 episodes: success 0.96, return 24.0, length 21.84
• local clip backbone-only result on same task: success 0.0, return 0.0
• local elastic reveal proxy iter6 result on same task: success 0.0, return 0.0
• local RVT frozen fixed-bounds result on same task: success 0.0, return 0.0

RVT overlap branch

• VLAarchtests/artifacts/reports/rvt_overlap_branch_20260330/summary.md
• VLAarchtests/artifacts/reports/rvt_overlap_branch_fixedbounds_20260330/summary.md

• frozen RVT stage1 train summary:
  • outputs/rlbench_rvt_branch/rlbench_subset3_backbone_only_rvt_100demo_frozen_seed17/summary.json
  • final train total 0.043179353826920445
  • final val total 0.039591669984665984
• frozen RVT overlap eval: mean success 0.0
• frozen fixed-bounds RVT overlap eval: mean success 0.0
• both branch gates:
  • local AnyBimanual overlap floor 0.16
  • stage2 run false

Dual-push non-privileged retarget branch

• VLAarchtests/artifacts/reports/dual_push_nonzero_branch_20260330/summary.md

• demo replay through absolute_action_from_delta:
  • reports/dual_push_nonzero_branch_20260330/demo_replay/replay_summary.json
  • mean success 0.8
  • mean return 0.8
• retargeted demo with checkpoint backbone retrieval and vision-only button localization:
  • reports/dual_push_nonzero_branch_20260330/retargeted_demo_backbone_vision_ep1/summary.json
  • mean success 1.0
  • mean return 1.0
• retargeted demo with checkpoint backbone retrieval and vision-only button localization:
  • reports/dual_push_nonzero_branch_20260330/retargeted_demo_backbone_vision_ep5/summary.json
  • mean success 1.0
  • mean return 1.0

Dual-push full-architecture hybrid branch

• VLAarchtests/artifacts/reports/dual_push_full_arch_hybrid_20260331/summary.md
• reports/dual_push_full_arch_probe_iter6_scene_ep1/summary.json
• reports/dual_push_full_arch_hybrid_iter6_backbone_ep1/summary.json

• elastic checkpoint retargeted-demo probe with scene retrieval and vision-only button localization:
  • 1 episode
  • mean success 1.0
  • mean return 1.0
  • steps 94
  • retrieved episode index 11
  • retrieval similarity 0.9998629689216614
• full-architecture hybrid eval with elastic controller checkpoint plus dual-push retrieval checkpoint:
  • 1 episode
  • mean success 1.0
  • mean return 1.0
  • steps 116
  • path recoveries 0
  • noop fallbacks 0
  • first selected mode residual::maintain_opening
  • last selected mode residual::base_action

Environment Recreation

• environment/setup_same_hardware.sh
• environment/runtime_env_vars.sh
• environment/reconstruct_anybimanual_overlap_replay.sh
• environment/hardware_snapshot.txt
• environment/env_list.txt
• environment/base_python.txt
• environment/base_pip_freeze.txt
• environment/rlbench_python.txt
• environment/rlbench_pip_freeze.txt

Notes On Result Presentation

• file paths
• exact commands
• exact numeric outputs
• exact partial status for in-flight runs