Add EgoTouch T-ReX editor replica checkpoint and strict loader

README.md

@@ -57,3 +57,7 @@ Its checkpoint uses `human_trex.full.v1`, not the legacy baseline format describ
 [model/tachin_adaptation_20260922](model/tachin_adaptation_20260922) contains validation-selected SmolVLA, VITRA, and HumanTReX checkpoints after Tachin adaptation.
 These are trained-parameter checkpoints, not standalone full models. The release documents their frozen initialization dependencies.
 No tactile editor or full test results are included. The existing three-domain and EgoTouch releases remain unchanged.
+
+## EgoTouch T-ReX editor replica
+
+[model/trex_editor_egotouch_replica_20260930](model/trex_editor_egotouch_replica_20260930) contains the validation-selected tactile editor from the September 30 video-generation replica. This is not the original paper-run checkpoint. It is separate from the full HumanTReX backbone. The package includes a strict loader and a verification record.

model/trex_editor_egotouch_replica_20260930/README.md

@@ -0,0 +1,58 @@
+# EgoTouch T-ReX editor — September 2026 replica
+
+This package contains the external tactile editor for our no-touch HumanTReX backbone.
+It is the September 30 replica for video generation, not the original paper-run checkpoint.
+The replica does not replace the original reported results.
+
+## Files
+
+- `best.pt`: the validation-selected editor, including its trained tactile trunk and training statistics.
+- `config.json`: the original run config. Local paths describe the original environment, not download requirements.
+- `load_editor.py`: a strict loader for the public editor package.
+- `verification.json`: checkpoint hash and CPU inference smoke results.
+
+The selected checkpoint is from epoch 5, step 20,870. The run trained for five epochs.
+The tactile CNN/GRU started from random initialization and trained with the editor.
+This is not the P1 retrieval encoder or the W3 editor. It uses no memory tokens or world-model losses.
+The backbone remained frozen through cached predictions.
+
+## Dependencies
+
+Backbone: [model/trex_full_egotouch](../trex_full_egotouch).
+This editor package does not contain the backbone, datasets, or caches.
+
+Editor implementation: [T-WLA-public/tactile_editor](https://github.com/kaichen-z/VLA-HAND/tree/d0731ac045298641e37251fd16cebd24e6f7b34c/tactile_editor).
+Install that package with `pip install -e ./tactile_editor` from the repository root.
+
+## Download and load
+
+```python
+from pathlib import Path
+import sys
+from huggingface_hub import snapshot_download
+
+root = Path(snapshot_download(
+    "MIT-Media-Lab/vla-hand-baselines",
+    allow_patterns=["model/trex_editor_egotouch_replica_20260930/*"],
+)) / "model/trex_editor_egotouch_replica_20260930"
+sys.path.insert(0, str(root))
+from load_editor import load_editor
+editor, statistics = load_editor(root / "best.pt", device="cuda")
+```
+
+## Input contract
+
+The editor accepts normalized base actions `[B,16,192]`, eight raw EgoTouch pressure frames `[B,8,2,21,21]`, and temporal validity `[B,8,2]`.
+Additional inputs are the edit mask `[B,16,192]` and plan-relative decision time `[B]` in seconds.
+The internal trunk applies the left-glove grid flip. Input histories must not apply that flip a second time.
+
+Normalize raw actions with `statistics['action_mean']` and `statistics['action_std']` before the editor call.
+Convert edited actions back to physical units with the same statistics.
+The editor returns `(normalized_actions, details)`.
+
+Offline replay uses 10 Hz plans, 80 Hz actions, and decisions at 0, 33.3, and 66.7 ms.
+Each revision starts from the immutable base plan. The mask excludes committed actions and padding dimensions 138:192.
+Evaluation executes the first eight actions with the latest available correction.
+
+The original config notes upstream split exposure. These are development results, not evidence of a fully isolated benchmark.
+The release smoke test checks strict loading, finite outputs, and unchanged masked actions. It is not a new full evaluation.

model/trex_editor_egotouch_replica_20260930/best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0567384be68a055610620ae7bdad97fbedf8318a03dca24f54f14331995ae6fd
+size 2461723

model/trex_editor_egotouch_replica_20260930/config.json

@@ -0,0 +1,56 @@
+{
+  "root": "/scratch/chonghej/VLA/video_regeneration_20260929/trex",
+  "variant": "trex",
+  "epochs": 5,
+  "micro_batch": 64,
+  "effective_batch": 256,
+  "workers": 4,
+  "seed": 20260921,
+  "smoke_steps": 0,
+  "resume": false,
+  "output_dir": "/scratch/chonghej/VLA/video_regeneration_20260929/trex/runs/trex",
+  "lambda_rgb": 0.0,
+  "lambda_tactile": 0.0,
+  "diagnostic_every": 0,
+  "smoke_full_weights": false,
+  "action_lr": 0.0001,
+  "encoder_lr": 0.0001,
+  "aux_delay_epochs": 0,
+  "aux_ramp_steps": 400,
+  "encoder_freeze_epochs": 0,
+  "roi_background": 0.1,
+  "wm_action_gradient": 0.0,
+  "validation_only": false,
+  "train_rows": 1068435,
+  "validation_rows": 60640,
+  "query_identity": {
+    "plans": 415624,
+    "edit_rows": 1068435,
+    "counts_by_tick": [
+      415624,
+      353877,
+      298934
+    ],
+    "query_sha256": "e4f4fb8e3e552fb8bfc2fd2bd18b99ba0f776c90f869821e0a5e350639050307",
+    "decision_ticks": [
+      0,
+      1,
+      2
+    ],
+    "action_hz": 80,
+    "replan_hz": 10,
+    "tactile_hz": 30,
+    "rgb_target_reference": "plan_start + [0.1,0.2] seconds",
+    "future_tactile_reference": "decision + 0.1 seconds",
+    "baseline_gt_max_difference": 0.0,
+    "baseline_mask_mismatch": 0,
+    "t0_touch_max_difference": 0.0,
+    "source_time_policy": "original source frames at 30 Hz; PCHIP pose labels at 80 Hz; no extrapolation; causal hold-last RGB for missing historical frames"
+  },
+  "action_scaling": "shared train-GT mean/std",
+  "world_weight_warmup": 400,
+  "seed_note": "same row shuffle and effective batch for runs with the same seed",
+  "test_exposure": "See upstream_split_audit.json; cross-backbone comparison remains a development result.",
+  "trainable_components": "standard editor including randomly initialized CNN/GRU tactile trunk; cached trex frozen; no memory tokens enter editor",
+  "pretrained_retrieval_encoder": false
+}

model/trex_editor_egotouch_replica_20260930/load_editor.py

@@ -0,0 +1,19 @@
+"""Load the EgoTouch T-ReX editor replica with the public editor package."""
+import torch
+from tactile_editor.direct_residual import DirectTactileResidualEditor
+
+
+def load_editor(checkpoint, device="cpu"):
+    # Load only checkpoints from a trusted source. This file contains metadata too.
+    saved = torch.load(checkpoint, map_location="cpu", weights_only=False)
+    if saved["config"]["variant"] != "trex":
+        raise ValueError("Expected a T-ReX editor checkpoint")
+    stats = saved["statistics"]
+    model = DirectTactileResidualEditor(
+        action_mean=torch.zeros(192), action_std=torch.ones(192),
+        pressure_masks=stats["pressure_masks"],
+        reference_pressure=stats["reference_pressure"],
+        mode="history", residual_bases=8, articulation_only=False,
+    )
+    model.load_state_dict(saved["model"], strict=True)
+    return model.to(device).eval(), stats

model/trex_editor_egotouch_replica_20260930/verification.json

@@ -0,0 +1,34 @@
+{
+  "status": "PASS",
+  "replica": true,
+  "original_paper_checkpoint": false,
+  "code_revision": "d0731ac045298641e37251fd16cebd24e6f7b34c",
+  "strict_load": true,
+  "synthetic_cpu_forward": true,
+  "finite_output": true,
+  "masked_actions_unchanged": true,
+  "shape": [
+    3,
+    16,
+    192
+  ],
+  "full_evaluation_rerun": false,
+  "files": {
+    "best.pt": {
+      "bytes": 2461723,
+      "sha256": "0567384be68a055610620ae7bdad97fbedf8318a03dca24f54f14331995ae6fd"
+    },
+    "config.json": {
+      "bytes": 1878,
+      "sha256": "eee545aba960b75083ddbbda044e81fbe3103afd174145bc93976748f4548ae7"
+    },
+    "README.md": {
+      "bytes": 2977,
+      "sha256": "af8b9d6ffb8fcf51f484bf62478aca7c96d450a1736e92ddf84b5fdae65f5e87"
+    },
+    "load_editor.py": {
+      "bytes": 886,
+      "sha256": "1c27e83657550f1d7ac4af52e348a9006bd80ce2c107b2d88b37b5de70253c13"
+    }
+  }
+}