Add binary/continuous gripper modes with DITA runtime 2.0.0

Update policy/processor configs, runtime wheel, documentation and real-data validation. Preserve model weights and LIBERO statistic tensors.

.gitattributes

@@ -34,3 +34,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 lerobot_policy_dita-1.0.5-py3-none-any.whl filter=lfs diff=lfs merge=lfs -text
+lerobot_policy_dita-2.0.0-py3-none-any.whl filter=lfs diff=lfs merge=lfs -text

GRIPPER_MODES.md

@@ -0,0 +1,122 @@
+# DITA gripper modes
+
+Runtime **2.0.0** supports `policy.gripper_mode=binary` and `continuous`. The mode is
+saved in `config.json` and both processor JSON files. The public LeRobot processor
+factory applies the active policy mode when loading pretrained templates, including
+CLI overrides. Conflicting explicit processor modes fail. New fine-tuning binds the
+target dataset statistics; resume keeps the saved statistics.
+
+The config schema replaces the serialized `action_normalization_mask` and
+`gripper_sign_flip` fields with `gripper_mode`. Old policy and processor configs are
+not supported. Use the updated Hub revisions and their 2.0.0 wheel together.
+
+| Contract | `binary` | `continuous` |
+| --- | --- | --- |
+| Target dataset | LIBERO signed commands | SO101 EEF gripper positions |
+| Training gripper | `-1=open, +1=closed` becomes `1=open, 0=closed` | Dataset `action.min/max` maps position to `[-1, 1]` |
+| Model output | Continuous diffusion prediction | Continuous diffusion prediction |
+| Policy output | `prediction > 0.5` becomes `+1`, otherwise `-1` | No threshold |
+| Postprocessor output | Sign flip restores `-1=open, +1=closed` | Inverse min/max mapping restores recorded position units |
+| Hub defaults | Spatial, Object, Goal, Long | DROID fine-tuning initialization |
+
+Both modes use target-dataset `action.q01/q99` for the first six pose dimensions.
+Continuous gripper additionally requires seven-element finite `action.min/max`.
+Gripper outputs are bounded to the recorded min/max; a constant gripper range maps
+to zero during training and restores the constant position during inference.
+
+The SO101 merged dataset has gripper min/max approximately `0.377/47.472` while its
+q01/q99 are approximately `1.227/28.432`. Using the full gripper min/max retains
+rare fully open positions that would exceed the diffusion sampler's normalized
+range with quantile scaling. Dataset extrema describe recorded commands, not new
+hardware calibration limits. Use the same robot calibration for recording and rollout.
+
+Binary mode accepts raw LIBERO signed labels. Already transformed RLDS `0/1`
+labels must first be converted to LIBERO commands (`1 - 2 * openness`) before using
+this processor. This restores the original DITA learning convention while keeping
+all converted LIBERO inference outputs unchanged.
+
+## Fine-tune
+
+Install the runtime bundled with the updated initialization:
+
+```bash
+pip install 'lerobot_policy_dita[training] @ https://huggingface.co/albus2024/dita_droid_pretrain/resolve/main/lerobot_policy_dita-2.0.0-py3-none-any.whl'
+```
+
+LIBERO:
+
+```bash
+lerobot-train \
+  --policy.path=albus2024/dita_droid_pretrain \
+  --policy.gripper_mode=binary \
+  --policy.device=cuda --policy.dtype=bfloat16 \
+  --policy.input_features=null --policy.output_features=null \
+  --policy.push_to_hub=false \
+  --dataset.repo_id=HuggingFaceVLA/libero \
+  --output_dir=outputs/train/dita_libero_binary
+```
+
+SO101 EEF, using the local merged dataset and front camera:
+
+```bash
+lerobot-train \
+  --policy.path=albus2024/dita_droid_pretrain \
+  --policy.gripper_mode=continuous \
+  --policy.device=cuda --policy.dtype=bfloat16 \
+  --policy.input_features='{"observation.images.image":{"type":"VISUAL","shape":[3,480,640]}}' \
+  --policy.output_features=null \
+  --rename_map='{"observation.images.front":"observation.images.image"}' \
+  --policy.push_to_hub=false \
+  --dataset.repo_id=albus2024/so101_merged \
+  --dataset.root=/home/lyanam/project/lerobot_real_robot/datasets/so101_merged \
+  --dataset.video_backend=pyav \
+  --output_dir=outputs/train/dita_so101_continuous
+```
+
+Keep `traj_length=11`, `chunk_size=10`, `n_action_steps=1`, and the existing camera
+and temporal contract. The DROID artifact is initialization without target-dataset
+statistics. Fine-tune with the chosen mode; changing modes after training does not
+convert learned gripper targets or align EEF coordinate frames.
+
+## EEF deployment
+
+Load the fine-tuned policy **and its saved processors** in `lerobot-rollout-eef`.
+After the policy postprocessor, the seven action channels must be ordered as:
+
+```text
+ee.x, ee.y, ee.z, ee.wx, ee.wy, ee.wz, ee.gripper_pos
+```
+
+`ee.gripper_pos` is already a continuous value in the recorded servo units. The
+EEF IK processor forwards it to `gripper.pos`; no additional threshold, sign flip,
+or multiplication by 100 is needed. SO101 data uses absolute position and rotvec
+orientation; preserve that action interpretation, robot calibration, and camera mapping.
+
+## Reproduce the real-data mechanics checks
+
+From the repository root, with an updated local DROID artifact:
+
+```bash
+CUDA_VISIBLE_DEVICES=0 python scripts/validate_gripper_modes.py \
+  --checkpoint /path/to/updated/dita_droid_pretrain \
+  --dataset-root /path/to/libero \
+  --dataset-repo-id HuggingFaceVLA/libero \
+  --mode binary --output workspace/validation/gripper_libero
+
+CUDA_VISIBLE_DEVICES=1 python scripts/validate_gripper_modes.py \
+  --checkpoint /path/to/updated/dita_droid_pretrain \
+  --dataset-root /home/lyanam/project/lerobot_real_robot/datasets/so101_merged \
+  --dataset-repo-id albus2024/so101_merged \
+  --camera-key observation.images.front \
+  --mode continuous --output workspace/validation/gripper_so101
+```
+
+The script invokes native `lerobot.scripts.lerobot_train` for ten optimizer steps,
+resumes for one more step, and checks finite weights/optimizer state, changed weights,
+saved statistics, real action encode/decode, `[4, 10, 7]` chunk inference, and
+`[4, 7]` online single-frame inference. It records exact commands, logs, and
+`verification.json`. `--verify-only` rechecks existing checkpoints without training.
+
+The 2026-09-14 checks passed on three episodes of each real dataset with BF16 and
+batch size four. They validate training and inference mechanics, not convergence,
+LIBERO task success, or a physical SO101 rollout.

README.md

@@ -38,7 +38,7 @@ This repository is a strict LeRobot conversion of `libero_long.pth` from the ori
 ## Install
 
 ```bash
-pip install "lerobot_policy_dita[libero] @ https://huggingface.co/albus2024/dita_long/resolve/main/lerobot_policy_dita-1.0.5-py3-none-any.whl"
+pip install "lerobot_policy_dita[libero] @ https://huggingface.co/albus2024/dita_long/resolve/main/lerobot_policy_dita-2.0.0-py3-none-any.whl"
 ```
 
 ## Evaluate
@@ -46,6 +46,7 @@ pip install "lerobot_policy_dita[libero] @ https://huggingface.co/albus2024/dita
 ```bash
 lerobot-eval \
   --policy.path=albus2024/dita_long \
+  --policy.gripper_mode=binary \
   --policy.device=cuda \
   --env.type=libero \
   --env.task=libero_10 \
@@ -64,14 +65,20 @@ The converter checks all parameter names and shapes, loads with `strict=True`, a
 saved artifact files. Run the project's parity and converted-checkpoint verification scripts
 separately before using the policy; conversion alone does not establish rollout parity.
 
-## Runtime 1.0.5 refresh
+## Runtime 2.0.0 gripper contract
 
-Use the bundled 1.0.5 wheel for this revision. The configuration now records
-`dtype=float16` to preserve the original CUDA AMP default; BF16 can be selected
-with `--policy.dtype=bfloat16`. The original model weights and processor files
-are unchanged. Older runtimes can use the previous revision recorded in
-`runtime_metadata.json`. Original conversion metadata and earlier reports are
-historical; `runtime_validation_report.json` records the checks run with this wheel.
+`gripper_mode` replaces the serialized action mask and gripper sign fields. Old
+policy/processor configs are not supported. Install the 2.0.0 wheel with this
+revision. See [gripper modes](GRIPPER_MODES.md) for training and SO101 EEF use.
 
-This refresh checks loading, processing and numerical execution. It does not
-repeat a full LIBERO benchmark or establish DROID task success.
+The four LIBERO models default to `binary`; the DROID initialization defaults to
+`continuous` and has no target-dataset statistics. For LIBERO fine-tuning from
+DROID, pass `--policy.gripper_mode=binary`. Both saved processors follow that
+override. Binary labels are -1=open/+1=closed; continuous gripper uses dataset
+min/max while pose uses quantiles.
+
+The runtime validation report covers real LIBERO and SO101 training (10 steps plus
+one resumed update), offline inference, wheel verification, four LIBERO checkpoint
+action parity, and three LIBERO Spatial simulator steps. These checks do not
+establish convergence, task success, or a physical SO101 rollout. Model weights
+and existing LIBERO statistics are unchanged in this update.

artifact_sha256.json

@@ -1,15 +1,15 @@
 {
-  "README.md": "29252d10abe939bfa7492b43ab54feaa2403c8115242ceda72db08a78331d29c",
-  "config.json": "752c8faff2deec9a8248d6c6ff49670a3ab77b667a581bf5396d4e6628b88441",
-  "conversion_metadata.json": "d15664a8af19e9072d0e581e334f8e8e77e2cc093ab5b5224e0947772616567d",
-  "lerobot_policy_dita-1.0.5-py3-none-any.whl": "8e4a6c5bf9d52a636751b97237305c13454d8091a9eb900fabcac875a7ae809e",
-  "local_verification.json": "25b91f455fb8708da85597e859a6cf98f237aaeec1d3720dd0a41c8b367b2b5d",
+  "GRIPPER_MODES.md": "92373ff4adbdb04cf5d5c415b9f56c29d768111a8fa712b669d2e0fb763ea5a4",
+  "README.md": "75327986b27402c095f05826c19a0814ab3eecf69b3bb4e1c574571b0be50f27",
+  "config.json": "9878d0fb0aa1ffd3fe7ba2a43c7225091fff7d801d51ca016f6420750ee6f820",
+  "conversion_metadata.json": "ee97a849125d84ca767c3ee9616af27896ac43fb1b0caed9fd608a770bbf7849",
+  "lerobot_policy_dita-2.0.0-py3-none-any.whl": "b948a1e1b2f2fb22ab1f952195ccae3006e08ad3f2695825533262296395ce01",
   "model.safetensors": "b111459ee0921631a9b5bbbb04713856081750d8e9f33664ee13889a41151be2",
-  "policy_postprocessor.json": "4f80b4d5c45a6f004a30021bbf17d8d6cbe621048ccd0364ca23cf496c8baac2",
+  "policy_postprocessor.json": "2668a59651f8fc638914065f71d7efebc48addca535891e38d75d4c2ba5e924c",
   "policy_postprocessor_step_0_dita_masked_unnormalizer_processor.safetensors": "e6bef89048da8a99aff85aa4263a02097af441185b534df37c453be302a949fe",
-  "policy_preprocessor.json": "3a61832dee2a66f9697dfc47eca57fb98a659c4c61c856c378f359074ca38aaa",
+  "policy_preprocessor.json": "ead05cc569fedb9f4ad5bc1338af09cd2174a1b83ac324be56f5c48049286a01",
   "policy_preprocessor_step_4_dita_masked_normalizer_processor.safetensors": "e6bef89048da8a99aff85aa4263a02097af441185b534df37c453be302a949fe",
-  "runtime_metadata.json": "5da64bf52df5c69f0e10c1b46e8a160ac1a6a7ab5c5c25096b685d84f7fda147",
-  "runtime_source_manifest.json": "b5a59da8d67fef584e1caa7ad11bc89a1621f57468c0df0663e78fee2b04ee46",
-  "runtime_validation_report.json": "07d1b3e42178e554791f2e9f9f099bd5b1fc474793b2d745c4f6577b601abe3b"
+  "runtime_metadata.json": "a2ccff24ca51dc35f8e5d56d61702bf2c508cc2b553e3274babae9ce1bfbfa03",
+  "runtime_source_manifest.json": "9da1a65a4edc8a1fe56a580fe4ace28358b42393fb8a6d0170e6bb6a2df9af04",
+  "runtime_validation_report.json": "1872e9a2560ff916960b55aa70b26f98ac328b2c1c16c7724b191bff69b81792"
 }

config.json

@@ -46,16 +46,6 @@
   "n_action_steps": 1,
   "history_frames": 2,
   "gripper_threshold": 0.5,
-  "gripper_sign_flip": -1.0,
-  "action_normalization_mask": [
-    true,
-    true,
-    true,
-    true,
-    true,
-    true,
-    false
-  ],
   "normalization_mapping": {
     "VISUAL": "IDENTITY",
     "ACTION": "QUANTILES",
@@ -69,5 +59,6 @@
   "optimizer_eps": 1e-08,
   "optimizer_weight_decay": 0.01,
   "optimizer_grad_clip_norm": 1.0,
-  "dtype": "float16"
+  "dtype": "float16",
+  "gripper_mode": "binary"
 }

conversion_metadata.json

@@ -46,5 +46,10 @@
     "policy_postprocessor_step_0_dita_masked_unnormalizer_processor.safetensors",
     "policy_preprocessor_step_4_dita_masked_normalizer_processor.safetensors"
   ],
-  "previous_hub_revision": "4fa5f721f7f16621361a8b23fa77cee1bc1c952f"
+  "previous_hub_revision": "4fa5f721f7f16621361a8b23fa77cee1bc1c952f",
+  "runtime_update": {
+    "version": "2.0.0",
+    "gripper_mode": "binary",
+    "metadata": "runtime_metadata.json"
+  }
 }

lerobot_policy_dita-2.0.0-py3-none-any.whl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b948a1e1b2f2fb22ab1f952195ccae3006e08ad3f2695825533262296395ce01
+size 129502

policy_postprocessor.json

@@ -18,16 +18,7 @@
           "ACTION": "QUANTILES",
           "LANGUAGE": "IDENTITY"
         },
-        "action_normalization_mask": [
-          true,
-          true,
-          true,
-          true,
-          true,
-          true,
-          false
-        ],
-        "gripper_sign_flip": -1.0
+        "gripper_mode": "binary"
       },
       "state_file": "policy_postprocessor_step_0_dita_masked_unnormalizer_processor.safetensors"
     },
@@ -39,4 +30,4 @@
       }
     }
   ]
-}
+}

policy_preprocessor.json

@@ -56,15 +56,7 @@
           "ACTION": "QUANTILES",
           "LANGUAGE": "IDENTITY"
         },
-        "action_normalization_mask": [
-          true,
-          true,
-          true,
-          true,
-          true,
-          true,
-          false
-        ]
+        "gripper_mode": "binary"
       },
       "state_file": "policy_preprocessor_step_4_dita_masked_normalizer_processor.safetensors"
     },
@@ -76,4 +68,4 @@
       }
     }
   ]
-}
+}

runtime_metadata.json

@@ -1,27 +1,23 @@
 {
-  "updated_at_utc": "2026-09-14T08:28:58.781286+00:00",
-  "runtime_version": "1.0.5",
-  "runtime_wheel": "lerobot_policy_dita-1.0.5-py3-none-any.whl",
-  "runtime_wheel_sha256": "8e4a6c5bf9d52a636751b97237305c13454d8091a9eb900fabcac875a7ae809e",
-  "source_git_head": "c97999bb495b3d58b105772690138a2cb7e8a709",
+  "updated_at_utc": "2026-09-14T10:39:32.234200+00:00",
+  "runtime_version": "2.0.0",
+  "minimum_runtime_version": "2.0.0",
+  "runtime_wheel": "lerobot_policy_dita-2.0.0-py3-none-any.whl",
+  "runtime_wheel_sha256": "b948a1e1b2f2fb22ab1f952195ccae3006e08ad3f2695825533262296395ce01",
+  "previous_hub_revision": "e222ad78a1905b39a38c31b169c62640d4666e4c",
+  "source_git_head": "7bc16687b141cdb7617e2a9dc6456e8ace1fedae",
   "source_git_dirty": true,
   "source_manifest": "runtime_source_manifest.json",
-  "previous_hub_revision": "2b76d79456232134252554fe19cc2c88de7bd199",
   "model_sha256": "b111459ee0921631a9b5bbbb04713856081750d8e9f33664ee13889a41151be2",
   "model_weights_changed": false,
-  "processor_files_changed": false,
+  "processor_configs_changed": true,
+  "processor_statistics_changed": false,
   "processor_sha256": {
+    "policy_postprocessor.json": "2668a59651f8fc638914065f71d7efebc48addca535891e38d75d4c2ba5e924c",
+    "policy_preprocessor.json": "ead05cc569fedb9f4ad5bc1338af09cd2174a1b83ac324be56f5c48049286a01",
     "policy_postprocessor_step_0_dita_masked_unnormalizer_processor.safetensors": "e6bef89048da8a99aff85aa4263a02097af441185b534df37c453be302a949fe",
-    "policy_preprocessor_step_4_dita_masked_normalizer_processor.safetensors": "e6bef89048da8a99aff85aa4263a02097af441185b534df37c453be302a949fe",
-    "policy_preprocessor.json": "3a61832dee2a66f9697dfc47eca57fb98a659c4c61c856c378f359074ca38aaa",
-    "policy_postprocessor.json": "4f80b4d5c45a6f004a30021bbf17d8d6cbe621048ccd0364ca23cf496c8baac2"
+    "policy_preprocessor_step_4_dita_masked_normalizer_processor.safetensors": "e6bef89048da8a99aff85aa4263a02097af441185b534df37c453be302a949fe"
   },
-  "configuration_change": {
-    "dtype": {
-      "before": "absent (CUDA AMP default float16)",
-      "after": "float16"
-    }
-  },
-  "minimum_runtime_version": "1.0.5",
+  "gripper_mode": "binary",
   "validation_report": "runtime_validation_report.json"
 }

runtime_source_manifest.json

@@ -1,10 +1,10 @@
 {
-  "wheel": "lerobot_policy_dita-1.0.5-py3-none-any.whl",
-  "sha256": "8e4a6c5bf9d52a636751b97237305c13454d8091a9eb900fabcac875a7ae809e",
+  "wheel": "lerobot_policy_dita-2.0.0-py3-none-any.whl",
+  "sha256": "b948a1e1b2f2fb22ab1f952195ccae3006e08ad3f2695825533262296395ce01",
   "source_python_files": 79,
   "all_source_files_match": true,
   "source_sha256": {
-    "lerobot_policy_dita/__init__.py": "b30cea9585424b15b93ddcf730d3b42d24372f78a063fc870af74b8763034b19",
+    "lerobot_policy_dita/__init__.py": "885069ffa087e7a89a2fb1461df108617e4632fe4ddfe7a3f62070b9b50c2601",
     "lerobot_policy_dita/_loading.py": "325e7c57a422f2cc5827c6d95ad19df4b51acae4af0a9b70b0c98a37852e6689",
     "lerobot_policy_dita/approximation/__init__.py": "b0646b441737d19c58ce395dd402fa0c1dcb0e8531e4df8d8be251de20b11d5e",
     "lerobot_policy_dita/approximation/adapters/__init__.py": "e0b5bf1a2a01a82af6b1ab03954aa039f396096b8b3a67d17483a7f054fd98c5",
@@ -33,18 +33,18 @@
     "lerobot_policy_dita/evaluation/static.py": "758ea97ef6b68c0a32197d962d9ff9af62b2aeb337c248bc72d8d960cdeba712",
     "lerobot_policy_dita/evaluation/trace.py": "a4399277a49e412589df69529d952e5931d87742f0a6dd16832d39cba113e1d3",
     "lerobot_policy_dita/policy/__init__.py": "9785f8e7279dc6760b4b7dbbe587ed6c6f467daa1dbbdf2ea095a645c1c52a57",
-    "lerobot_policy_dita/policy/_lerobot_compat.py": "84e49b026ed2b9cff8f079ff0839fef528a25e02e986aa4bcf79906c949e9b36",
+    "lerobot_policy_dita/policy/_lerobot_compat.py": "d3434f59edf9d1dbb5616b874d0d6babf67a254f0b48196559e06d0dae1c6214",
     "lerobot_policy_dita/policy/approximation/__init__.py": "0632173bd9a1595a72c9741e6bb5538f644c0f448c8caca606542ae3d2cb3b5a",
     "lerobot_policy_dita/policy/approximation/adapter.py": "d603f444c2237a76a23f44dd8b2df734b66525d3522695f200478b3c055152a4",
     "lerobot_policy_dita/policy/approximation/configuration.py": "8fa1f0f24624352b691d73c71f9fc75c0196c8f81e888e6b77a9f11eba7926af",
     "lerobot_policy_dita/policy/approximation/ddim.py": "d32a16f85aa670bd07375ce2691498c47b9f9e700160fde12eea33e6a56f09fa",
     "lerobot_policy_dita/policy/attention.py": "3253866b5ea0bdf24a85b2047d813573e24588a4294adc9b72d01488d236909e",
-    "lerobot_policy_dita/policy/configuration_dita.py": "3817b74d2d30366d097e537897a12a20f0ce08d57c389de10d0458893870bee4",
+    "lerobot_policy_dita/policy/configuration_dita.py": "379bbbb0124e445d3930e527c5e8706451f6c16d76775f04fffdb97230761f3b",
     "lerobot_policy_dita/policy/evaluation/__init__.py": "8c0ca0bbd994637ce95307daec7647082fdb3e8e982073ab6c60ba7f9a16745f",
     "lerobot_policy_dita/policy/evaluation/adapter.py": "fdbfadfbfa13772eef20a513c74a06cda0e6cea456a4b2197cb314af42c2dbec",
     "lerobot_policy_dita/policy/evaluation/static.py": "4517d7af032dee8ef0f73217a0dda023cc48d8fc6a39624b26c9647d71b0f169",
-    "lerobot_policy_dita/policy/modeling_dita.py": "dcf0766a3c92988bd38c55dab8c902a5b8cd97443fab00fde4189aa93c7931c0",
-    "lerobot_policy_dita/policy/processor_dita.py": "e0c14c6b2a217e28ca6cf7e9654a24134471d8efe09acd7481358be1f3fe0707",
+    "lerobot_policy_dita/policy/modeling_dita.py": "1fb98b081deb3c47f51180e31874abdf342b2fb9e5b8d5d8d7bef5e60a0c7a1d",
+    "lerobot_policy_dita/policy/processor_dita.py": "3173c16c7be81bdb5f86cce305eee8f96a16235acfb45586f874e80324f8257b",
     "lerobot_policy_dita/policy/quantization/__init__.py": "8b130b24c8c2a6e5b585c1e5c80e37ba317578ded5bfc624b68a9548d82fcd67",
     "lerobot_policy_dita/policy/quantization/adapter.py": "1633d4eb613972afe14d0de3138ec45391764702449e1b7f6312cfeb209d2068",
     "lerobot_policy_dita/policy/quantization/attention.py": "e54d5105cb5930050b268d53e098bfcde6303908884f3d909e83bb50e740ec5e",

runtime_validation_report.json

@@ -1,33 +1,470 @@
 {
-  "source": "albus2024/dita_long",
-  "dataset_suite": "libero_10_no_noops",
-  "strict_from_pretrained": true,
-  "processor_stats_roundtrip": true,
-  "model_roundtrip_max_abs_diff": 0.0,
-  "inference_steps": 3,
-  "actions_finite": true,
-  "action_shape": [
-    1,
-    7
-  ],
-  "gripper_values": [
-    -1.0,
-    -1.0,
-    -1.0
-  ],
-  "history_lengths": [
-    1,
-    2,
-    2
-  ],
-  "result": "passed",
-  "runtime_version": "1.0.5",
-  "runtime_from_bundled_wheel": true,
-  "dtype_cli_overrides": [
-    "float32",
-    "float16",
-    "bfloat16"
-  ],
-  "weights_and_processor_files_unchanged": true,
-  "full_task_benchmark_run": false
+  "runtime_version": "2.0.0",
+  "unit_tests_passed": 305,
+  "libero_four_checkpoint_real_frame_parity": true,
+  "libero_spatial_simulator_steps": 3,
+  "hardware_rollout_tested": false,
+  "task_success_evaluated": false,
+  "real_data_validation": {
+    "libero": {
+      "mode": "binary",
+      "dataset_repo_id": "HuggingFaceVLA/libero",
+      "episodes": [
+        0,
+        1,
+        2
+      ],
+      "frames": 843,
+      "training_steps": 10,
+      "resume_steps": 1,
+      "batch_size": 4,
+      "losses": [
+        0.201,
+        0.139,
+        0.288,
+        0.059,
+        0.156,
+        0.183,
+        0.152,
+        0.092,
+        0.241,
+        0.215,
+        0.441
+      ],
+      "gradient_norms": [
+        6.592,
+        4.542,
+        7.649,
+        2.729,
+        4.631,
+        4.119,
+        4.009,
+        3.905,
+        4.769,
+        7.744,
+        13.014
+      ],
+      "reloaded_loss": 0.10520374029874802,
+      "changed_weight_tensors": 198,
+      "all_model_and_optimizer_tensors_finite": true,
+      "optimizer_steps": [
+        11.0
+      ],
+      "processor_stats_preserved_on_resume": true,
+      "real_action_encode_decode_roundtrip": true,
+      "gripper_bounds_source": "LIBERO signed commands",
+      "dataset_gripper_min_max": [
+        -1.0,
+        1.0
+      ],
+      "bound_gripper_min_max": [
+        -1.0,
+        1.0
+      ],
+      "inference_shape": [
+        4,
+        10,
+        7
+      ],
+      "selected_action_shape": [
+        4,
+        7
+      ],
+      "inference_gripper_values": [
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "runtime_version": "2.0.0",
+      "versions": {
+        "lerobot": "0.6.1",
+        "torch": "2.9.1",
+        "accelerate": "1.14.0",
+        "diffusers": "0.39.0"
+      },
+      "scope": "Real-data training, resume and offline inference; no task success or hardware rollout claim"
+    },
+    "so101": {
+      "mode": "continuous",
+      "dataset_repo_id": "albus2024/so101_merged",
+      "episodes": [
+        0,
+        1,
+        2
+      ],
+      "frames": 1341,
+      "training_steps": 10,
+      "resume_steps": 1,
+      "batch_size": 4,
+      "losses": [
+        0.188,
+        0.379,
+        0.258,
+        0.078,
+        0.139,
+        0.227,
+        0.131,
+        0.091,
+        0.319,
+        0.227,
+        0.544
+      ],
+      "gradient_norms": [
+        6.483,
+        9.196,
+        7.543,
+        2.857,
+        4.061,
+        4.045,
+        3.402,
+        3.099,
+        6.443,
+        4.464,
+        8.722
+      ],
+      "reloaded_loss": 0.32114681601524353,
+      "changed_weight_tensors": 198,
+      "all_model_and_optimizer_tensors_finite": true,
+      "optimizer_steps": [
+        11.0
+      ],
+      "processor_stats_preserved_on_resume": true,
+      "real_action_encode_decode_roundtrip": true,
+      "gripper_bounds_source": "action.min/max",
+      "dataset_gripper_min_max": [
+        0.37735849618911743,
+        47.47169876098633
+      ],
+      "bound_gripper_min_max": [
+        0.37735849618911743,
+        47.47169876098633
+      ],
+      "inference_shape": [
+        4,
+        10,
+        7
+      ],
+      "selected_action_shape": [
+        4,
+        7
+      ],
+      "inference_gripper_values": [
+        2.7805802822113037,
+        4.978194713592529,
+        6.185774326324463,
+        6.798575401306152,
+        5.922640800476074,
+        12.053372383117676,
+        11.256832122802734,
+        19.64849853515625,
+        10.308959007263184,
+        5.831273078918457,
+        16.023744583129883,
+        11.27537727355957,
+        7.672808647155762,
+        8.728656768798828,
+        7.07811164855957,
+        7.678109645843506,
+        8.474404335021973,
+        10.399279594421387,
+        5.7238264083862305,
+        1.8922457695007324,
+        19.525482177734375,
+        13.28436279296875,
+        12.212488174438477,
+        7.542376518249512,
+        7.725221633911133,
+        7.23825216293335,
+        9.160019874572754,
+        9.65340518951416,
+        6.483657360076904,
+        9.079004287719727,
+        2.6967198848724365,
+        0.8949151635169983,
+        2.467010021209717,
+        1.1595251560211182,
+        5.166204929351807,
+        5.179502010345459,
+        4.2831854820251465,
+        0.5067476630210876,
+        0.44492819905281067,
+        1.639739751815796
+      ],
+      "runtime_version": "2.0.0",
+      "versions": {
+        "lerobot": "0.6.1",
+        "torch": "2.9.1",
+        "accelerate": "1.14.0",
+        "diffusers": "0.39.0"
+      },
+      "scope": "Real-data training, resume and offline inference; no task success or hardware rollout claim"
+    },
+    "wheel_libero": {
+      "mode": "binary",
+      "dataset_repo_id": "HuggingFaceVLA/libero",
+      "episodes": [
+        0,
+        1,
+        2
+      ],
+      "frames": 843,
+      "training_steps": 10,
+      "resume_steps": 1,
+      "batch_size": 4,
+      "losses": [
+        0.201,
+        0.139,
+        0.288,
+        0.059,
+        0.156,
+        0.183,
+        0.152,
+        0.092,
+        0.241,
+        0.215,
+        0.441
+      ],
+      "gradient_norms": [
+        6.592,
+        4.542,
+        7.649,
+        2.729,
+        4.631,
+        4.119,
+        4.009,
+        3.905,
+        4.769,
+        7.744,
+        13.014
+      ],
+      "reloaded_loss": 0.29260557889938354,
+      "changed_weight_tensors": 198,
+      "all_model_and_optimizer_tensors_finite": true,
+      "optimizer_steps": [
+        11.0
+      ],
+      "processor_stats_preserved_on_resume": true,
+      "real_action_encode_decode_roundtrip": true,
+      "gripper_bounds_source": "LIBERO signed commands",
+      "dataset_gripper_min_max": [
+        -1.0,
+        1.0
+      ],
+      "bound_gripper_min_max": [
+        -1.0,
+        1.0
+      ],
+      "inference_shape": [
+        4,
+        10,
+        7
+      ],
+      "selected_action_shape": [
+        4,
+        7
+      ],
+      "inference_gripper_values": [
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        -1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0,
+        1.0
+      ],
+      "runtime_version": "2.0.0",
+      "versions": {
+        "lerobot": "0.6.1",
+        "torch": "2.9.1",
+        "accelerate": "1.14.0",
+        "diffusers": "0.39.0"
+      },
+      "scope": "Real-data training, resume and offline inference; no task success or hardware rollout claim"
+    },
+    "wheel_so101": {
+      "mode": "continuous",
+      "dataset_repo_id": "albus2024/so101_merged",
+      "episodes": [
+        0,
+        1,
+        2
+      ],
+      "frames": 1341,
+      "training_steps": 10,
+      "resume_steps": 1,
+      "batch_size": 4,
+      "losses": [
+        0.188,
+        0.379,
+        0.258,
+        0.078,
+        0.139,
+        0.227,
+        0.131,
+        0.091,
+        0.319,
+        0.227,
+        0.544
+      ],
+      "gradient_norms": [
+        6.483,
+        9.196,
+        7.543,
+        2.857,
+        4.061,
+        4.045,
+        3.402,
+        3.099,
+        6.443,
+        4.464,
+        8.722
+      ],
+      "reloaded_loss": 0.15004707872867584,
+      "changed_weight_tensors": 198,
+      "all_model_and_optimizer_tensors_finite": true,
+      "optimizer_steps": [
+        11.0
+      ],
+      "processor_stats_preserved_on_resume": true,
+      "real_action_encode_decode_roundtrip": true,
+      "gripper_bounds_source": "action.min/max",
+      "dataset_gripper_min_max": [
+        0.37735849618911743,
+        47.47169876098633
+      ],
+      "bound_gripper_min_max": [
+        0.37735849618911743,
+        47.47169876098633
+      ],
+      "inference_shape": [
+        4,
+        10,
+        7
+      ],
+      "selected_action_shape": [
+        4,
+        7
+      ],
+      "inference_gripper_values": [
+        2.7805802822113037,
+        4.978194713592529,
+        6.185774326324463,
+        6.798575401306152,
+        5.922640800476074,
+        12.053372383117676,
+        11.256832122802734,
+        19.64849853515625,
+        10.308959007263184,
+        5.831273078918457,
+        16.023744583129883,
+        11.27537727355957,
+        7.672808647155762,
+        8.728656768798828,
+        7.07811164855957,
+        7.678109645843506,
+        8.474404335021973,
+        10.399279594421387,
+        5.7238264083862305,
+        1.8922457695007324,
+        19.525482177734375,
+        13.28436279296875,
+        12.212488174438477,
+        7.542376518249512,
+        7.725221633911133,
+        7.23825216293335,
+        9.160019874572754,
+        9.65340518951416,
+        6.483657360076904,
+        9.079004287719727,
+        2.6967198848724365,
+        0.8949151635169983,
+        2.467010021209717,
+        1.1595251560211182,
+        5.166204929351807,
+        5.179502010345459,
+        4.2831854820251465,
+        0.5067476630210876,
+        0.44492819905281067,
+        1.639739751815796
+      ],
+      "runtime_version": "2.0.0",
+      "versions": {
+        "lerobot": "0.6.1",
+        "torch": "2.9.1",
+        "accelerate": "1.14.0",
+        "diffusers": "0.39.0"
+      },
+      "scope": "Real-data training, resume and offline inference; no task success or hardware rollout claim"
+    }
+  }
 }