configs: organize Cartesian experiments by task

docs/ARC_CLOCK_CONTRACT.md

@@ -4,6 +4,12 @@ This contract replaces the historical Cartesian mean/wide/stacked/clock/E1
 families. It builds on PR #263 and preserves the local first-row-delay
 implementation. It does not change the separate planar or Libero codecs.
 
+New run presets are grouped by task under `hydra_configs/experiment/`:
+`organize_stationary`, `fold_and_stack_towels`, `wireless_earbuds`, and
+`multitask`, each with `arc_vel`, `arc_dur`, and `baseline`.
+See the [experiment menu](../egomimic/hydra_configs/experiment/README.md)
+for models, frozen data pools, budgets, retained variants and migration.
+
 ```yaml
 _target_: egomimic.rldb.zarr.arc_length_tokenizer.TokenizeBimanualArcLengthCartesian
 clock_mode: velocity                 # velocity | duration

docs/experiment_config_migration.json

@@ -0,0 +1,32 @@
+{
+  "abc_arc/cotrain/organize_clock_duration_full_separation": "organize_stationary/arc_dur",
+  "abc_arc/cotrain/organize_clock_duration_translation_rotation": "organize_stationary/arc_dur",
+  "abc_arc/cotrain/organize_clock_velocity_full_separation": "organize_stationary/arc_vel",
+  "abc_arc/cotrain/organize_clock_velocity_translation_rotation": "organize_stationary/arc_vel",
+  "abc_arc/cotrain/organize_rl2_elmo_hpt180_cotrain_baseline": "organize_stationary/variants/organize_rl2_elmo_hpt180_cotrain_baseline",
+  "abc_arc/cotrain/organize_rl2_elmo_hpt180_cotrain_multistream": "organize_stationary/variants/organize_rl2_elmo_hpt180_cotrain_multistream",
+  "abc_arc/cotrain/organize_rl2_elmo_hpt300_cotrain_baseline": "organize_stationary/baseline",
+  "abc_arc/cotrain/organize_rl2_elmo_hpt300_cotrain_multistream": "organize_stationary/_base/arc",
+  "abc_arc/cotrain/organize_rl2_elmo_hpt300_cotrain_multistream_logclock": "organize_stationary/arc_dur",
+  "abc_arc/cotrain/organize_rl2_elmo_hpt300_cotrain_multistream_perwaypoint": "organize_stationary/arc_vel",
+  "abc_arc/robot_bc/abc_multitask4_abcdit_baseline_visual_openloop": "multitask/variants/abc_multitask4_abcdit_baseline_visual_openloop",
+  "abc_arc/robot_bc/abc_multitask4_abcdit_hybrid_visual_openloop": "multitask/variants/abc_multitask4_abcdit_hybrid_visual_openloop",
+  "abc_arc/robot_bc/abc_multitask4_hpt300_baseline_visual_openloop": "multitask/baseline",
+  "abc_arc/robot_bc/abc_multitask4_hpt300_hybrid_visual_openloop": "multitask/arc_vel",
+  "abc_arc/robot_bc/abc_rl2_towels_abcdit_baseline_visual_openloop": "fold_and_stack_towels/variants/abc_rl2_towels_abcdit_baseline_visual_openloop",
+  "abc_arc/robot_bc/abc_rl2_towels_abcdit_multistream_logclock_visual_openloop": "fold_and_stack_towels/variants/abc_rl2_towels_abcdit_multistream_logclock_visual_openloop",
+  "abc_arc/robot_bc/abc_rl2_towels_abcdit_multistream_visual_openloop": "fold_and_stack_towels/variants/abc_rl2_towels_abcdit_multistream_visual_openloop",
+  "abc_arc/robot_bc/abc_rl2_towels_hpt180_multistream_visual_openloop": "fold_and_stack_towels/variants/abc_rl2_towels_hpt180_multistream_visual_openloop",
+  "abc_arc/robot_bc/abc_towels_hpt180_baseline_visual_openloop": "fold_and_stack_towels/variants/abc_towels_hpt180_baseline_visual_openloop",
+  "abc_arc/robot_bc/abc_towels_hpt180_hybrid_visual_openloop": "fold_and_stack_towels/variants/abc_towels_hpt180_hybrid_visual_openloop",
+  "abc_arc/robot_bc/abc_visual_hpt300_base": "_shared/visual_hpt300",
+  "abc_arc/robot_bc/stationery_rl2_hpt300_visual_baseline_openloop": "organize_stationary/variants/stationery_rl2_hpt300_visual_baseline_openloop",
+  "abc_arc/robot_bc/stationery_rl2_hpt300_visual_hybrid_openloop": "organize_stationary/variants/stationery_rl2_hpt300_visual_hybrid_openloop",
+  "yam_arc_grid/_base/earbuds_updated_hpt300": "_shared/frozen_yam/earbuds_updated_hpt300",
+  "yam_arc_grid/_base/stattempo": "_shared/frozen_yam/stattempo",
+  "yam_arc_grid/_base/stattempo_openloop": "_shared/frozen_yam/stattempo_openloop",
+  "yam_arc_grid/_base/towels_hpt300": "_shared/frozen_yam/towels_hpt300",
+  "yam_arc_grid/earbuds_updated_time_hpt300": "wireless_earbuds/baseline",
+  "yam_arc_grid/towels_new290plus74_time_hpt300": "fold_and_stack_towels/variants/new290plus74_baseline",
+  "yam_arc_grid/towels_since0924plus74_time_hpt300": "fold_and_stack_towels/baseline"
+}

egomimic/hydra_configs/data/abc_arc/abc_visual.yaml

@@ -29,7 +29,8 @@ train_datasets:
         min_distance_unit: ${abc.arc_distance}
         rotation_distance_unit: ${abc.arc_rotation_distance}
         resampled_vector_length: ${abc.arc_waypoints}
-        clock_mode: velocity
+        clock_mode: ${abc.arc_clock_mode}
+        streams: ${abc.arc_streams}
     filters:
       _target_: egomimic.rldb.filters.DatasetFilter
       filter_lambdas:

egomimic/hydra_configs/experiment/README.md

@@ -0,0 +1,64 @@
+# Task-first experiments
+
+The Cartesian YAM/ABC experiment menu is:
+
+```text
+experiment/
+  organize_stationary/   {baseline,arc_vel,arc_dur}.yaml
+  fold_and_stack_towels/  {baseline,arc_vel,arc_dur}.yaml
+  wireless_earbuds/       {baseline,arc_vel,arc_dur}.yaml
+  multitask/             {baseline,arc_vel,arc_dur}.yaml
+  _shared/               shared defaults, not standalone runs
+```
+
+Select one with `+experiment=organize_stationary/arc_vel` (or `arc_dur`
+or `baseline`). Cluster resources still come from the explicit
+`hydra/launcher=...` selection and require a separate launch preflight.
+
+| Folder | Default model / data | ARC M / D / R |
+| --- | --- | --- |
+| organize_stationary | HPT300, RL2 Elmo robot + human organize stationery | 100 / 0.42 m / 24° |
+| fold_and_stack_towels | HPT300, frozen RL2 `since0924plus74` pool | 200 / 0.33 m / 100° |
+| wireless_earbuds | HPT300, frozen RL2 `wireless_earbuds_updated` pool | 200 / 0.33 m / 100° |
+| multitask | HPT300, existing ABC four-task pool | 100 / 0.81 m / 24° |
+
+These are explicit recipe defaults, not a new distance/rotation calibration.
+The four-task pool is towels, fake fruits, stationery and earbuds, **not**
+the separate seven-task ABC-130k proposal.
+Each task's three entries use exactly the same filters, episode lists and split.
+Baseline geometry is 100 × 14 time-indexed actions.
+
+ARC uses the canonical [clock contract](../../../docs/ARC_CLOCK_CONTRACT.md)
+(repository path `docs/ARC_CLOCK_CONTRACT.md`): velocity or log **interval**
+seconds, first-row start delay, first-moving-stream termination.
+It is not E1 or cumulative absolute seconds. Full separation is the default.
+Use `abc.arc_streams=translation_rotation` for four clocks; token width
+automatically changes from M × 20 to M × 18, including the model and decoder.
+No tokenizer or detokenizer algorithm is changed by this YAML reorganization.
+
+The frozen towel/earbud ARC entries are **new canonical-codec recipes** built
+on their retained baseline network/optimizer/pool. They read 400 native
+frames, use the shared open-loop evaluator, and require fresh normalization
+and training. They do not reproduce retired E1 ARC checkpoints.
+The baselines retain their original evaluator/training schedule.
+
+## Retained variants
+
+Distinct existing HPT180, robot-only, ABC-DiT, ABC-only/ABC+RL2 towel,
+and `new290plus74` pool choices live under each task's `variants/`.
+They retain their model/data settings and are not interchangeable aliases for
+the primary menu. Other task/model families (PushT, Libero, human-only,
+Ravens, etc.) are untouched.
+
+Duplicate organize clock-layout leaves were consolidated: choose
+`organize_stationary/arc_vel` or `arc_dur` and set `abc.arc_streams` instead
+of choosing a second file. Distinct old experiment paths are mapped in
+`docs/experiment_config_migration.json`; the two retired
+`*_translation_rotation` aliases additionally require
+`abc.arc_streams=translation_rotation`.
+
+Old CLI paths under `abc_arc/{robot_bc,cotrain}` and `yam_arc_grid` are removed,
+not silently redirected. Original jobs/checkpoints must keep their recorded
+source/config/normalization. No cluster jobs are migrated by changing these
+files. Historical E1 recipe-generation scripts are not a supported way to
+regenerate this task menu.

egomimic/hydra_configs/experiment/_shared/clock_settings.yaml

@@ -0,0 +1,24 @@
+# @package _global_
+# Shared Cartesian clock contract; task recipes own D, R, M and the data pool.
+abc:
+  arc_clock_mode: velocity
+  arc_streams: full_separation
+  arc_stream_dims:
+    full_separation: 20
+    translation_rotation: 18
+  arc_token_dim: ${abc.arc_stream_dims.${abc.arc_streams}}
+  arc_token_rows: ${abc.arc_waypoints}
+run_provenance:
+  action_contract:
+    representation: ${abc.action_mode}
+    arc_chunking_mode: ${abc.arc_chunking_mode}
+    translation_distance_m: ${abc.arc_distance}
+    rotation_distance_radians: ${abc.arc_rotation_distance}
+    waypoints: ${abc.arc_waypoints}
+    clock_mode: ${abc.arc_clock_mode}
+    streams: ${abc.arc_streams}
+    timing_layout: first_delay_incoming_v1
+    hold_window_frames: 400
+    gripper_stream_mode: ${oc.select:abc.arc_gripper_stream_mode,events}
+    gripper_distance_unit: ${oc.select:abc.arc_gripper_distance,null}
+    control_dt: ${evaluator.control_dt}

egomimic/hydra_configs/experiment/yam_arc_grid/_base/earbuds_updated_hpt300.yaml → egomimic/hydra_configs/experiment/_shared/frozen_yam/earbuds_updated_hpt300.yaml

@@ -1,6 +1,6 @@
 # @package _global_
 defaults:
-- /experiment/yam_arc_grid/_base/towels_hpt300
+- /experiment/_shared/frozen_yam/towels_hpt300
 - override /data: abc_arc/earbuds_updated
 - _self_
 e1:

egomimic/hydra_configs/experiment/yam_arc_grid/_base/stattempo_openloop.yaml → egomimic/hydra_configs/experiment/_shared/frozen_yam/stattempo_openloop.yaml

@@ -3,7 +3,7 @@
 # validation, for robot-only BC and (via _base/stattempo_aria) Aria cotrain.
 # Keep evaluator selection in defaults so teacher-forced kwargs cannot leak in.
 defaults:
-- /experiment/yam_arc_grid/_base/stattempo
+- /experiment/_shared/frozen_yam/stattempo
 - override /evaluator: eval_open_loop_sim
 - _self_
 trainer:

egomimic/hydra_configs/experiment/yam_arc_grid/_base/towels_hpt300.yaml → egomimic/hydra_configs/experiment/_shared/frozen_yam/towels_hpt300.yaml

@@ -3,7 +3,7 @@
 # the same training recipe and tempo evaluator as the stationery base; leaves pick the frozen pool
 # (towels_pool + data override), the codec (e1.variant) and the run identity.
 defaults:
-- /experiment/yam_arc_grid/_base/stattempo
+- /experiment/_shared/frozen_yam/stattempo
 - override /model: e1/hpt300_flow_wrists_time
 - override /data: abc_arc/towels_since0924plus74_20261015
 - _self_

egomimic/hydra_configs/experiment/_shared/frozen_yam_arc.yaml

@@ -0,0 +1,93 @@
+# @package _global_
+# Canonical ARC over the retained frozen YAM baseline graph and episode lists.
+defaults:
+  - /experiment/_shared/clock_settings
+  - /model_contract/abc_visual@model
+  - override /evaluator: eval_open_loop_sim
+  - _self_
+abc:
+  action_mode: hybrid_arc_tokenizer_cartesian
+  arc_chunking_mode: multistream
+  arc_distance: 0.33
+  arc_rotation_distance: 1.7453292519943295
+  arc_waypoints: 200
+  arc_source_buffer_frames: 400
+  arc_gripper_stream_mode: events
+  arc_gripper_distance: null
+e1:
+  variant: arc
+  yam_source_frames: ${abc.arc_source_buffer_frames}
+  M: ${abc.arc_waypoints}
+  D: ${abc.arc_distance}
+arc_tokenizer:
+  min_distance_unit: ${abc.arc_distance}
+  resampled_vector_length: ${abc.arc_waypoints}
+hpt:
+  action_horizon: ${abc.arc_token_rows}
+  token_dim: ${abc.arc_token_dim}
+model:
+  action_token_dim: ${abc.arc_token_dim}
+  action_horizon: ${abc.arc_token_rows}
+  native_action_horizon: ${abc.arc_token_rows}
+  native_action_dim: ${abc.arc_token_dim}
+  inference: ${model.inference_contracts.hybrid_arc_tokenizer_cartesian}
+  data_requirements:
+    preprocessing:
+      '7':
+        key_map:
+          _target_: egomimic.rldb.embodiment.bimanual_arc.get_keymap
+          horizon: 200
+          embodiment: yam
+          drop_wrist_images: false
+          yam_source_frames: ${abc.arc_source_buffer_frames}
+        transforms:
+          _target_: egomimic.rldb.embodiment.bimanual_arc.get_transform_list
+          variant: arc
+          chunk_length: 200
+          time_rows: ${abc.arc_source_buffer_frames}
+          min_distance_unit: ${abc.arc_distance}
+          rotation_distance_unit: ${abc.arc_rotation_distance}
+          resampled_vector_length: ${abc.arc_waypoints}
+          stride: 1
+          rotation_mode: euler
+          embodiment: yam
+          yam_source_frames: ${abc.arc_source_buffer_frames}
+          clock_mode: ${abc.arc_clock_mode}
+          streams: ${abc.arc_streams}
+          gripper_stream_mode: ${abc.arc_gripper_stream_mode}
+        source_fps: 30.0
+  graph:
+    noising:
+      action_horizon: ${abc.arc_token_rows}
+    sampler:
+      action_horizon: ${abc.arc_token_rows}
+      model:
+        act_seq: ${abc.arc_token_rows}
+data:
+  train_datasets:
+    yam_bimanual:
+      resolver:
+        key_map: ${model.data_requirements.preprocessing.7.key_map}
+        transform_list: ${model.data_requirements.preprocessing.7.transforms}
+  valid_datasets:
+    yam_bimanual:
+      resolver:
+        key_map: ${model.data_requirements.preprocessing.7.key_map}
+        transform_list: ${model.data_requirements.preprocessing.7.transforms}
+trainer:
+  limit_val_batches: 1.0
+evaluator:
+  action_mode: arc
+  ground_truth_action_key: actions_time
+  execute_fraction: 0.3
+  arc_execution_cap_mode: waypoints
+  arc_chunking_mode: ${abc.arc_chunking_mode}
+  min_distance_unit: ${abc.arc_distance}
+  rotation_distance_unit: ${abc.arc_rotation_distance}
+  resampled_vector_length: ${abc.arc_waypoints}
+  clock_mode: ${abc.arc_clock_mode}
+  streams: ${abc.arc_streams}
+logger:
+  wandb:
+    id: ${name}_s42
+    tags: [rl2, hpt300, "${abc.arc_clock_mode}", "${abc.arc_streams}", open_loop, video]

egomimic/hydra_configs/experiment/abc_arc/robot_bc/abc_visual_hpt300_base.yaml → egomimic/hydra_configs/experiment/_shared/visual_hpt300.yaml

@@ -1,5 +1,6 @@
 # @package _global_
 defaults:
+- /experiment/_shared/clock_settings
 - override /model: abc_arc/hpt_yam_visual
 - override /data: abc_arc/abc_visual
 - override /evaluator: eval_open_loop_sim
@@ -30,26 +31,8 @@ abc:
   arc_distance: 0.4
   arc_rotation_distance: null
   arc_waypoints: 100
-  arc_token_rows: 100
-  arc_token_dim: 20
-  arc_streams: full_separation
-  arc_clock_mode: velocity
   arc_chunking_mode: multistream
   task_predicate: ???
-run_provenance:
-  action_contract:
-    representation: ${abc.action_mode}
-    arc_chunking_mode: ${abc.arc_chunking_mode}
-    translation_distance_m: ${abc.arc_distance}
-    rotation_distance_radians: ${abc.arc_rotation_distance}
-    waypoints: ${abc.arc_waypoints}
-    clock_mode: ${abc.arc_clock_mode}
-    streams: ${abc.arc_streams}
-    timing_layout: first_delay_incoming_v1
-    hold_window_frames: 400
-    gripper_stream_mode: ${oc.select:abc.arc_gripper_stream_mode,events}
-    gripper_distance_unit: ${oc.select:abc.arc_gripper_distance,null}
-    control_dt: ${evaluator.control_dt}
 trainer:
   max_steps: 240000
   max_epochs: -1
@@ -74,7 +57,7 @@ evaluator:
   min_distance_unit: ${abc.arc_distance}
   rotation_distance_unit: ${abc.arc_rotation_distance}
   resampled_vector_length: ${abc.arc_waypoints}
-  clock_mode: velocity
+  clock_mode: ${abc.arc_clock_mode}
   limit_val_episodes: 16
   requires_ordered_validation: true
 logger:

egomimic/hydra_configs/experiment/abc_arc/README.md

@@ -1,20 +1,14 @@
-# Supported visual ABC / human BC experiments
+# Human-only BC and migrated task recipes
 
-Launch from this experiment group, not from retired model/data templates:
+Robot and cotrain presets now live in the four task folders documented in
+[the experiment menu](../README.md). Distinct HPT180, ABC-DiT and data-pool
+variants are preserved under those folders; duplicate clock-layout recipes
+are replaced by the `abc.arc_streams` override.
 
-| Group | Dataset | Paired recipes |
-| --- | --- | --- |
-| `robot_bc` | ABC stationery slice + RL2 updated organize-stationery; RL2-only validation | `stationery_rl2_hpt300_visual_{baseline,hybrid}_openloop` |
-| `robot_bc` | ABC fold and stack the towels | `abc_towels_hpt180_{baseline,hybrid}_visual_openloop` |
-| `robot_bc` | ABC four-task multitask | `abc_multitask4_hpt300_{baseline,hybrid}_visual_openloop` |
-| `human_bc` | MECKA folding clothes, 40-hour subset | `mecka_fold_clothes_40h_human_visual_{baseline,hybrid}_openloop` |
-
-Brace notation denotes separate YAML files, not Hydra syntax.
-
-`robot_bc/abc_visual_hpt300_base` is a shared template requiring a task
-predicate. Cotraining has four explicit
-`organize_clock_{velocity,duration}_{full_separation,translation_rotation}`
-leaves. See [the clock contract](../../../../docs/ARC_CLOCK_CONTRACT.md).
+This folder retains the independent MECKA human-only folding experiments:
+`human_bc/mecka_fold_clothes_40h_human_visual_{baseline,hybrid}_openloop`.
+Brace notation denotes two YAML files, not Hydra syntax.
+See [the clock contract](../../../../docs/ARC_CLOCK_CONTRACT.md).
 
 Select `abc.arc_clock_mode=velocity|duration` and
 `abc.arc_streams=full_separation|translation_rotation`, with token dimension