File size: 13,894 Bytes
bfb9665 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 | # Change And Test Log
This file records the main code changes and executed test commands copied into this repo. Result statements below are raw command outcomes only.
## Previous Repo Work Included Here
Copied from `history/VLAarchtests_previous_README.md`:
- core model, memory, planner, and dataset changes under:
- `VLAarchtests/code/reveal_vla_bimanual/models/`
- `VLAarchtests/code/reveal_vla_bimanual/train/losses.py`
- `VLAarchtests/code/reveal_vla_bimanual/sim_reveal/`
- `VLAarchtests/code/reveal_vla_bimanual/sim_rlbench/dataset.py`
- training and eval paths under:
- `VLAarchtests/code/reveal_vla_bimanual/train/`
- `VLAarchtests/code/reveal_vla_bimanual/eval/`
- earlier test suite under:
- `VLAarchtests/tests/`
## Current Session File Changes
### Core reveal/proxy path
- `VLAarchtests/code/reveal_vla_bimanual/models/policy.py`
- `VLAarchtests/code/reveal_vla_bimanual/models/action_decoder.py`
- `VLAarchtests/code/reveal_vla_bimanual/models/backbones.py`
- `VLAarchtests/code/reveal_vla_bimanual/models/rvt_backbone.py`
- `VLAarchtests/code/reveal_vla_bimanual/train/losses.py`
- `VLAarchtests/code/reveal_vla_bimanual/train/run_rlbench_experiment.py`
- `VLAarchtests/code/reveal_vla_bimanual/eval/run_reveal_benchmark.py`
- `VLAarchtests/code/reveal_vla_bimanual/eval/summarize_anybimanual_overlap_eval.py`
- `VLAarchtests/code/reveal_vla_bimanual/eval/summarize_rvt_overlap_branch.py`
- `VLAarchtests/code/reveal_vla_bimanual/eval/compose_task_routed_proxy_summary.py`
- `VLAarchtests/code/reveal_vla_bimanual/eval/run_proposal_alignment_diagnostics.py`
- `VLAarchtests/code/reveal_vla_bimanual/eval/run_rlbench_knn_eval.py`
- `VLAarchtests/code/reveal_vla_bimanual/eval/run_rlbench_knn_task_sweep.py`
- `VLAarchtests/code/reveal_vla_bimanual/eval/run_rlbench_dual_push_retargeted_demo_eval.py`
- `VLAarchtests/code/reveal_vla_bimanual/eval/run_rlbench_dual_push_full_arch_hybrid_eval.py`
- `VLAarchtests/code/reveal_vla_bimanual/eval/dual_push_retarget_utils.py`
- `VLAarchtests/code/reveal_vla_bimanual/eval/dual_push_full_arch_utils.py`
- `VLAarchtests/code/reveal_vla_bimanual/sim_reveal/build_task_specialized_episode_specs.py`
- `VLAarchtests/code/reveal_vla_bimanual/sim_reveal/procedural_envs.py`
- `VLAarchtests/code/reveal_vla_bimanual/sim_rlbench/task_resolver.py`
### Training/eval wrappers and configs
- `VLAarchtests/code/reveal_vla_bimanual/scripts/run_task_routed_proxy_eval.sh`
- `VLAarchtests/code/reveal_vla_bimanual/scripts/run_bag_selector_iter9.sh`
- `VLAarchtests/code/reveal_vla_bimanual/scripts/run_anybimanual_subset3_overlap_train.sh`
- `VLAarchtests/code/reveal_vla_bimanual/scripts/run_anybimanual_subset3_overlap_eval.sh`
- `VLAarchtests/code/reveal_vla_bimanual/scripts/run_rvt_overlap_branch.sh`
- `VLAarchtests/code/reveal_vla_bimanual/scripts/run_dual_push_retargeted_demo_eval.sh`
- `VLAarchtests/code/reveal_vla_bimanual/scripts/run_dual_push_full_arch_hybrid_eval.sh`
- `VLAarchtests/code/reveal_vla_bimanual/train/configs/proxy_interaction_r3d_stage3_clip_rgbd_handoff_compact_phase_v7_selector_finetune_iter6.yaml`
- `VLAarchtests/code/reveal_vla_bimanual/train/configs/proxy_interaction_r3d_stage3_clip_rgbd_handoff_compact_phase_v7_selector_finetune_iter7.yaml`
- `VLAarchtests/code/reveal_vla_bimanual/train/configs/proxy_interaction_r3d_stage3_clip_rgbd_handoff_compact_phase_v7_selector_finetune_iter8.yaml`
- `VLAarchtests/code/reveal_vla_bimanual/train/configs/proxy_interaction_r3d_stage3_clip_rgbd_handoff_compact_phase_v7_selector_finetune_iter9_bag.yaml`
- `VLAarchtests/code/reveal_vla_bimanual/train/configs/rlbench_subset3_backbone_only_clip_100demo_fair_step1_full.yaml`
- `VLAarchtests/code/reveal_vla_bimanual/train/configs/rlbench_subset3_backbone_only_rvt_100demo_frozen_seed17.yaml`
- `VLAarchtests/code/reveal_vla_bimanual/train/configs/rlbench_subset3_backbone_only_rvt_100demo_unfreeze_top2_seed17.yaml`
- `VLAarchtests/code/reveal_vla_bimanual/train/configs/rlbench_dual_push_backbone_only_clip_finetune_seed17.yaml`
- `VLAarchtests/code/reveal_vla_bimanual/train/configs/rlbench_dual_push_backbone_only_clip_finetune_weighted_seed17.yaml`
- `VLAarchtests/code/reveal_vla_bimanual/train/configs/rlbench_dual_push_backbone_only_clip_chunk8_weighted_seed17.yaml`
- `environment/reconstruct_anybimanual_overlap_replay.sh`
### Test additions or updates
- `VLAarchtests/tests/test_eval_toggle_paths_work.py`
- `VLAarchtests/tests/test_task_routed_model_eval.py`
- `VLAarchtests/tests/test_anybimanual_resume_logic.py`
- `VLAarchtests/tests/test_anybimanual_overlap_eval_summary.py`
- `VLAarchtests/tests/test_candidate_ranking_loss.py`
- `VLAarchtests/tests/test_compose_task_routed_proxy_summary.py`
- `VLAarchtests/tests/test_build_task_specialized_episode_specs.py`
- `VLAarchtests/tests/test_proposal_mode_names_label_base_action.py`
- `VLAarchtests/tests/test_proxy_scripted_bench.py`
- `VLAarchtests/tests/test_rvt_backbone_forward.py`
- `VLAarchtests/tests/test_rlbench_dataset_rgbd_geometry.py`
- `VLAarchtests/tests/test_rlbench_init_checkpoint.py`
- `VLAarchtests/tests/test_rlbench_pickle_bootstrap.py`
- `VLAarchtests/tests/test_rlbench_task_resolver_aliases.py`
- `VLAarchtests/tests/test_summarize_rvt_overlap_branch.py`
- `VLAarchtests/tests/test_dual_push_retarget_utils.py`
- `VLAarchtests/tests/test_dual_push_full_arch_utils.py`
### Third-party baseline path changes
- `third_party/AnyBimanual/third_party/YARR/yarr/runners/offline_train_runner.py`
- `third_party/AnyBimanual/third_party/YARR/yarr/runners/weight_init_utils.py`
- `third_party/AnyBimanual/agents/peract_bc/launch_utils.py`
- `third_party/AnyBimanual/agents/peract_bc/qattention_peract_bc_agent.py`
- `third_party/AnyBimanual/agents/peract_bimanual/qattention_peract_bc_agent.py`
## Current Session Test Commands
Executed commands recorded in the workspace:
- `python -m py_compile /workspace/VLAarchtests/code/reveal_vla_bimanual/models/action_decoder.py /workspace/VLAarchtests/tests/test_proposal_mode_names_label_base_action.py`
- `PYTHONPATH=/workspace/VLAarchtests/code/reveal_vla_bimanual pytest -q /workspace/VLAarchtests/tests/test_proposal_mode_names_label_base_action.py /workspace/VLAarchtests/tests/test_candidate_ranking_loss.py /workspace/VLAarchtests/tests/test_compose_task_routed_proxy_summary.py /workspace/VLAarchtests/tests/test_build_task_specialized_episode_specs.py`
- result: `11 passed`
- `pytest -q /workspace/VLAarchtests/tests/test_anybimanual_overlap_eval_summary.py`
- result: `2 passed`
- `pytest -q /workspace/VLAarchtests/tests/test_task_routed_model_eval.py /workspace/VLAarchtests/tests/test_eval_toggle_paths_work.py`
- result: `4 passed`
- `pytest -q /workspace/VLAarchtests/tests/test_rvt_backbone_forward.py /workspace/VLAarchtests/tests/test_rlbench_dataset_rgbd_geometry.py /workspace/VLAarchtests/tests/test_eval_toggle_paths_work.py /workspace/VLAarchtests/tests/test_rlbench_init_checkpoint.py /workspace/VLAarchtests/tests/test_rlbench_pickle_bootstrap.py /workspace/VLAarchtests/tests/test_rlbench_task_resolver_aliases.py /workspace/VLAarchtests/tests/test_summarize_rvt_overlap_branch.py`
- result: `passed`
- `pytest -q /workspace/VLAarchtests/tests/test_build_task_specialized_episode_specs.py /workspace/VLAarchtests/tests/test_candidate_ranking_loss.py /workspace/VLAarchtests/tests/test_compose_task_routed_proxy_summary.py`
- result: `10 passed`
- `pytest -q /workspace/VLAarchtests/tests/test_dual_push_retarget_utils.py /workspace/VLAarchtests/tests/test_rlbench_knn_eval_scene_kwargs.py`
- result: `passed`
- `pytest -q /workspace/VLAarchtests/tests/test_dual_push_retarget_utils.py`
- result: `6 passed`
- `pytest -q /workspace/VLAarchtests/tests/test_dual_push_retarget_utils.py /workspace/VLAarchtests/tests/test_dual_push_full_arch_utils.py`
- result: `9 passed`
- `bash -n /workspace/VLAarchtests/code/reveal_vla_bimanual/scripts/run_bag_selector_iter9.sh`
- `bash -n /workspace/VLAarchtests/code/reveal_vla_bimanual/scripts/run_task_routed_proxy_eval.sh`
- `bash -n /workspace/VLAarchtests/code/reveal_vla_bimanual/scripts/run_anybimanual_subset3_overlap_train.sh`
- `bash -n /workspace/VLAarchtests/code/reveal_vla_bimanual/scripts/run_rvt_overlap_branch.sh`
- `bash -n /workspace/VLAarchtests/code/reveal_vla_bimanual/scripts/run_dual_push_retargeted_demo_eval.sh`
- `bash -n /workspace/VLAarchtests/code/reveal_vla_bimanual/scripts/run_dual_push_full_arch_hybrid_eval.sh`
- `PYTHONPATH=/workspace/third_party/AnyBimanual/third_party/YARR pytest -q /workspace/VLAarchtests/tests/test_anybimanual_resume_logic.py`
- result: `4 passed`
- `python -m py_compile /workspace/VLAarchtests/code/reveal_vla_bimanual/models/rvt_backbone.py /workspace/VLAarchtests/code/reveal_vla_bimanual/train/run_rlbench_experiment.py /workspace/VLAarchtests/code/reveal_vla_bimanual/sim_rlbench/dataset.py /workspace/VLAarchtests/code/reveal_vla_bimanual/sim_rlbench/task_resolver.py /workspace/VLAarchtests/code/reveal_vla_bimanual/eval/summarize_rvt_overlap_branch.py`
- result: `passed`
- `python -m py_compile /workspace/VLAarchtests/code/reveal_vla_bimanual/eval/dual_push_retarget_utils.py /workspace/VLAarchtests/code/reveal_vla_bimanual/eval/run_rlbench_dual_push_retargeted_demo_eval.py`
- result: `passed`
- `python -m py_compile /workspace/VLAarchtests/code/reveal_vla_bimanual/eval/dual_push_retarget_utils.py /workspace/VLAarchtests/code/reveal_vla_bimanual/eval/dual_push_full_arch_utils.py /workspace/VLAarchtests/code/reveal_vla_bimanual/eval/run_rlbench_dual_push_full_arch_hybrid_eval.py`
- result: `passed`
## Current Session Generated Reports
Current-session report roots staged in this repo:
- `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/anybimanual_overlap_baseline_20260330/resume1000_summary/`
- `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/`
## HF Packaging Notes
Raw packaging changes applied to the staged HF export:
- `baselines/AnyBimanual_overlap_replay/multi/` was reshaped from one flat directory into shard subdirectories:
- `00000-04999/`
- `05000-09999/`
- `10000-14999/`
- file count after reshape: `14034`
- reconstruction helper added at:
- `environment/reconstruct_anybimanual_overlap_replay.sh`
- exact rejected Hub error before reshape:
- `Your push was rejected because it contains too many files per directory. Each directory in your git repo can only contain up to 10000 files. Offending directories: /baselines/AnyBimanual_overlap_replay/multi/`
## Current Session Logs
Main logs staged in this repo:
- `reports/anybimanual_subset3_overlap_smoke200_fixpretrain_nowandb3_train.log`
- `reports/anybimanual_subset3_overlap_smoke200_fixpretrain_nowandb3_train_presavefix.log`
- `reports/anybimanual_subset3_overlap_resume1000_eval.log`
- `reports/anybimanual_subset3_overlap_resume1000_summary.log`
- `reports/task_routed_proxy_v1_rerun.log`
- `reports/run_bag_selector_iter9_prebuild.log`
- `reports/anybimanual_release_subset3_eval_ep5.log`
- `reports/rvt_overlap_branch_fixedbounds_20260330_chain.sh`
- `reports/dual_push_full_arch_hybrid_iter6_scene_ep5.log`
- `reports/dual_push_full_arch_hybrid_iter6_backbone_ep2_r005.log`
## Official Overlap Eval Final Raw Outputs
Sources:
- `reports/anybimanual_subset3_overlap_resume1000_eval.log`
- `VLAarchtests/artifacts/reports/anybimanual_overlap_baseline_20260330/resume1000_summary/summary.json`
Raw values:
- step `1000`
- local mean success `0.16`
- `coordinated_push_box`: success `0.0`, return `0.0`
- `coordinated_lift_ball`: success `0.0`, return `0.0`
- `dual_push_buttons`: success `0.48`, return `12.0`
## General-Task Anchor Raw Outputs
Sources:
- `VLAarchtests/artifacts/reports/general_task_anchor_20260330_dual_push_buttons/summary.json`
Raw values:
- 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: success `0.0`, return `0.0`
- local elastic reveal proxy iter6 result: success `0.0`, return `0.0`
- local RVT frozen fixed-bounds result: success `0.0`, return `0.0`
## Dual-Push Branch Raw Outputs
Sources:
- `VLAarchtests/artifacts/reports/dual_push_nonzero_branch_20260330/summary.md`
- `VLAarchtests/artifacts/reports/dual_push_full_arch_hybrid_20260331/summary.md`
Raw values:
- demo replay through `absolute_action_from_delta`: mean success `0.8`, mean return `0.8`
- retargeted demo with checkpoint backbone retrieval and vision-only button localization, `5` episodes: mean success `1.0`, mean return `1.0`
- elastic checkpoint retargeted-demo probe with scene retrieval and vision-only button localization, `1` episode: mean success `1.0`, mean return `1.0`
- 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`
|