Add files using upload-large-folder tool

README.md

@@ -0,0 +1,66 @@
+---
+license: other
+license_name: slim-license
+license_link: https://github.com/kzz1031/SLIM/blob/main/LICENSE
+library_name: slim-policy
+tags:
+  - robotics
+  - robot-manipulation
+  - flow-matching
+  - self-supervised-learning
+  - libero
+---
+
+# SLIM for LIBERO
+
+This repository contains the released SLIM Stage 2 policy checkpoint for
+LIBERO and LIBERO-Plus. SLIM is a compact latent interaction policy for robot
+manipulation.
+
+## Checkpoint
+
+- Stage 1: action-grounded masked trajectory prediction on LIBERO all+90
+- Stage 1 objective: IDM:FDM = 0.125:1 for 3 epochs
+- Stage 2: flow-matching policy training on LIBERO all for 40 epochs
+- Stage 1 and Stage 2 video backend: `torchvision_av`
+- Stage 1 EMA: enabled, momentum 0.999
+- Stage 2 EMA: disabled
+- Action horizon and execution chunk: 8
+- Image size: 224 x 224, agent and wrist views
+- State/action dimensions: 7/7
+
+The checkpoint is a plain PyTorch `state_dict` and loads directly with
+[SLIM](https://github.com/kzz1031/SLIM).
+
+## Results
+
+| Benchmark | Coverage | Score |
+| --- | ---: | ---: |
+| LIBERO | 2,000 / 2,000 | 97.50% |
+| LIBERO-Plus | 10,030 / 10,030 | 77.45% |
+
+LIBERO suite scores are 94.40% (LIBERO-10), 99.40% (Spatial), 99.40%
+(Object), and 96.80% (Goal). The complete LIBERO-Plus suite/category reports
+are included under `evaluation/`.
+
+## Usage
+
+Install SLIM and configure the DINOv2 and T5 paths as described in the SLIM
+README. Then run a policy server from the SLIM repository root:
+
+```bash
+python -m slim.serving.server \
+  --checkpoint /path/to/SLIM-LIBERO/checkpoints/epoch_40_pytorch_model.pt \
+  --port 10093 \
+  --bf16
+```
+
+The checkpoint requires the included `config.yaml` and `action_stats.json` to
+remain in the repository root. See `checkpoint_manifest.json` for hashes and
+the exact release revision.
+
+## Limitations
+
+This checkpoint is intended for research evaluation in LIBERO-compatible
+simulation environments. It should not be deployed on physical robots without
+task-specific safety validation and action-bound checks.

action_stats.json

@@ -0,0 +1,34 @@
+{
+  "q01": [
+    -0.7044642567634583,
+    -0.8008928298950195,
+    -0.9375,
+    -0.11464285850524902,
+    -0.1639285683631897,
+    -0.2239285707473755,
+    0.0
+  ],
+  "q99": [
+    0.9375,
+    0.8678571581840515,
+    0.9375,
+    0.13178572058677673,
+    0.19285714626312256,
+    0.335357129573822,
+    1.0
+  ],
+  "normalization": "q01_q99",
+  "action_dim": 7,
+  "action_normalized_dims": [
+    0,
+    1,
+    2,
+    3,
+    4,
+    5
+  ],
+  "action_indices": null,
+  "action_key": "action",
+  "state_key": "observation.state",
+  "note": "q01/q99 are length action_dim; only action_normalized_dims are normalized"
+}

checkpoint_manifest.json

@@ -0,0 +1,29 @@
+{
+  "format": "SLIM native PyTorch state_dict",
+  "model": "SLIM",
+  "benchmark": "LIBERO and LIBERO-Plus",
+  "checkpoint_stage": 2,
+  "checkpoint_epoch": 40,
+  "action_horizon": 8,
+  "execution_chunk": 8,
+  "state_dim": 7,
+  "action_dim": 7,
+  "checkpoint_sha256": "e8e6a600603c30087a2ceb9d4c158be1cf5e4ff9dbcf9b892d9d17df00dcbd85",
+  "config_sha256": "bc7f45c401c8852a2aa5ddb75d47b3d1f20249f6e303a7ccb350e5de920f5671",
+  "action_stats_sha256": "0c02c05ab6a3008ba67cd3b4e3e23680588bcec15203747f68bc5636a530854b",
+  "reference_evaluation": {
+    "libero": {
+      "episodes": 2000,
+      "successes": 1950,
+      "success_rate": 0.975
+    },
+    "libero_plus": {
+      "cases": 10030,
+      "successes": 7768,
+      "success_rate": 0.7745
+    }
+  },
+  "slim_repository": "https://github.com/kzz1031/SLIM",
+  "slim_commit": "51bcc75ce659fe5eeb275900bdf9ba2b48991cdf",
+  "strict_load_verified": true
+}

checkpoints/epoch_40_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8e6a600603c30087a2ceb9d4c158be1cf5e4ff9dbcf9b892d9d17df00dcbd85
+size 944561675

config.yaml

@@ -0,0 +1,100 @@
+seed: 42
+run:
+  name: libero_all_stage2_policy_h8_40ep_seed42_from_idm0125_s1v_s2v_noema
+  root: checkpoints/stage2
+  timestamp: false
+model:
+  name: SLIM
+  vision_encoder:
+    backbone_name: dinov2_vitb14
+    model_path: ${oc.env:DINOV2_MODEL_DIR}
+    image_size: 224
+    freeze_backbone: false
+    num_image_views: 2
+    vision_condition_mode: dense_patch
+    image_augment:
+      enabled: false
+  ema:
+    enabled: false
+    momentum: 0.999
+  action_dim: 7
+  state_dim: 7
+  action_horizon: 8
+  use_language_condition: true
+  language_encoder_path: ${oc.env:T5_MODEL_DIR}
+  language_cache: ${oc.env:SLIM_LANGUAGE_CACHE,null}
+  max_language_tokens: 32
+  language_embedding_dim: 512
+  inference_steps: 4
+  repeated_diffusion_steps: 4
+  decoder_hidden_dim: 1024
+  transformer:
+    hidden_dim: 768
+    num_layers: 16
+    num_heads: 12
+    ffn_ratio: 4.0
+    dropout: 0.0
+    num_action_register_tokens: 4
+    num_future_tokens: 512
+    use_state_condition: true
+    max_state_tokens: 1100
+    max_action_tokens: 16
+    policy_loss_weight: 1.0
+    idm_loss_weight: 0.0
+    fdm_loss_weight: 0.0
+    future_loss_type: norm_l1
+data:
+  sources:
+  - root_dir: ${oc.env:LIBERO_DATA_ROOT}
+    data_mix: libero_all
+    video_keys:
+    - observation.images.image
+    - observation.images.wrist_image
+    weight: 1.0
+  action_normalization: q01_q99
+  action_stats_dir: ${oc.env:SLIM_CACHE_DIR,.cache/slim}/action_stats/libero_all
+  per_device_batch_size: 16
+  episode_shuffle: true
+  eval_batch_size: 8
+  val_ratio: 0.0
+  eval_monitor_episode_ratio: 0.05
+  split_seed: 42
+  num_workers: 16
+  eval_num_workers: 8
+  prefetch_factor: 2
+  persistent_workers: true
+  dataloader_timeout_s: 300
+  video_backend: torchvision_av
+training:
+  stage: 2
+  objective: policy
+  max_epochs: 40
+  max_train_steps: 100000
+  num_warmup_steps: 5000
+  save_interval: 5000
+  eval_interval: 100
+  logging_frequency: 10
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 1
+  learning_rate:
+    base: 2.5e-05
+    action_model: 0.0001
+    vision_encoder: 1.0e-05
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 0.01
+logging:
+  trackers:
+  - jsonl
+  - wandb
+  project: slim
+  entity: null
+resume_skip_prefixes:
+- ema_vision_encoder.
+- _ema_fp32_

evaluation/libero/summary.txt

@@ -0,0 +1,5 @@
+libero_10: 472/500 = 94.40%
+libero_spatial: 497/500 = 99.40%
+libero_object: 497/500 = 99.40%
+libero_goal: 484/500 = 96.80%
+overall: 1950/2000 = 97.50%

evaluation/libero_plus/libero_10/summary_metrics.txt

@@ -0,0 +1,27 @@
+LIBERO-Plus evaluation summary (from episode markers)
+model: libero_all_stage2_policy_h8_40ep_seed42_from_idm0125_s1v_s2v_noema_epoch_40_pytorch_model.pt
+coverage: 2519/2519 (100.00%)
+rollouts:  74.79% (1884/2519)
+
+=== By suite ===
+libero_10   74.79% (1884/2519)
+
+=== By category ===
+Robot Initial States    43.00% (169/393)
+Camera Viewpoints       57.76% (242/419)
+Sensor Noise            78.84% (354/449)
+Light Conditions        81.75% (224/274)
+Background Textures     85.47% (247/289)
+Objects Layout          91.67% (286/312)
+Language Instructions   94.52% (362/383)
+
+=== By difficulty ===
+5   53.76% (379/705)
+4   73.42% (384/523)
+3   79.75% (386/484)
+2   89.56% (403/450)
+1   93.00% (332/357)
+
+=== Leaderboard row ===
+Model	Camera	Robot	Language	Light	Background	Noise	Layout	Total
+libero_all_stage2_policy_h8_40ep_seed42_from_idm0125_s1v_s2v_noema_epoch_40_pytorch_model.pt	57.76	43.00	94.52	81.75	85.47	78.84	91.67	74.79

evaluation/libero_plus/libero_goal/summary_metrics.txt

@@ -0,0 +1,28 @@
+LIBERO-Plus evaluation summary (from episode markers)
+model: libero_all_stage2_policy_h8_40ep_seed42_from_idm0125_s1v_s2v_noema_epoch_40_pytorch_model.pt
+coverage: 2591/2591 (100.00%)
+rollouts:  62.22% (1612/2591)
+
+=== By suite ===
+libero_goal   62.22% (1612/2591)
+
+=== By category ===
+Robot Initial States    20.29% (83/409)
+Camera Viewpoints       49.26% (201/408)
+Language Instructions   59.27% (243/410)
+Objects Layout          64.71% (275/425)
+Sensor Noise            74.67% (283/379)
+Background Textures     91.81% (258/281)
+Light Conditions        96.42% (269/279)
+
+=== By difficulty ===
+5      34.95% (202/578)
+4      54.22% (244/450)
+3      63.00% (298/473)
+2      74.83% (428/572)
+1      80.86% (321/397)
+None   98.35% (119/121)
+
+=== Leaderboard row ===
+Model	Camera	Robot	Language	Light	Background	Noise	Layout	Total
+libero_all_stage2_policy_h8_40ep_seed42_from_idm0125_s1v_s2v_noema_epoch_40_pytorch_model.pt	49.26	20.29	59.27	96.42	91.81	74.67	64.71	62.22

evaluation/libero_plus/libero_object/summary_metrics.txt

@@ -0,0 +1,27 @@
+LIBERO-Plus evaluation summary (from episode markers)
+model: libero_all_stage2_policy_h8_40ep_seed42_from_idm0125_s1v_s2v_noema_epoch_40_pytorch_model.pt
+coverage: 2518/2518 (100.00%)
+rollouts:  83.28% (2097/2518)
+
+=== By suite ===
+libero_object   83.28% (2097/2518)
+
+=== By category ===
+Robot Initial States    35.43% (141/398)
+Camera Viewpoints       83.33% (330/396)
+Objects Layout          83.62% (337/403)
+Background Textures     94.76% (235/248)
+Sensor Noise            95.50% (403/422)
+Language Instructions  100.00% (354/354)
+Light Conditions       100.00% (297/297)
+
+=== By difficulty ===
+5   69.29% (422/609)
+4   77.75% (374/481)
+3   85.21% (432/507)
+2   92.56% (473/511)
+1   96.59% (396/410)
+
+=== Leaderboard row ===
+Model	Camera	Robot	Language	Light	Background	Noise	Layout	Total
+libero_all_stage2_policy_h8_40ep_seed42_from_idm0125_s1v_s2v_noema_epoch_40_pytorch_model.pt	83.33	35.43	100.00	100.00	94.76	95.50	83.62	83.28

evaluation/libero_plus/libero_spatial/summary_metrics.txt

@@ -0,0 +1,27 @@
+LIBERO-Plus evaluation summary (from episode markers)
+model: libero_all_stage2_policy_h8_40ep_seed42_from_idm0125_s1v_s2v_noema_epoch_40_pytorch_model.pt
+coverage: 2402/2402 (100.00%)
+rollouts:  90.55% (2175/2402)
+
+=== By suite ===
+libero_spatial   90.55% (2175/2402)
+
+=== By category ===
+Robot Initial States    51.14% (179/350)
+Camera Viewpoints       95.21% (358/376)
+Sensor Noise            96.30% (338/351)
+Objects Layout          96.36% (371/385)
+Background Textures     96.90% (250/258)
+Language Instructions   99.23% (387/390)
+Light Conditions       100.00% (292/292)
+
+=== By difficulty ===
+5   68.06% (130/191)
+4   88.43% (382/432)
+3   90.00% (567/630)
+2   94.02% (629/669)
+1   97.29% (467/480)
+
+=== Leaderboard row ===
+Model	Camera	Robot	Language	Light	Background	Noise	Layout	Total
+libero_all_stage2_policy_h8_40ep_seed42_from_idm0125_s1v_s2v_noema_epoch_40_pytorch_model.pt	95.21	51.14	99.23	100.00	96.90	96.30	96.36	90.55

evaluation/libero_plus/summary.txt

@@ -0,0 +1,31 @@
+LIBERO-Plus evaluation summary (from episode markers)
+model: libero_all_stage2_policy_h8_40ep_seed42_from_idm0125_s1v_s2v_noema_epoch_40_pytorch_model.pt
+coverage: 10030/10030 (100.00%)
+rollouts:  77.45% (7768/10030)
+
+=== By suite ===
+libero_goal      62.22% (1612/2591)
+libero_10        74.79% (1884/2519)
+libero_object    83.28% (2097/2518)
+libero_spatial   90.55% (2175/2402)
+
+=== By category ===
+Robot Initial States    36.90% (572/1550)
+Camera Viewpoints       70.73% (1131/1599)
+Objects Layout          83.21% (1269/1525)
+Sensor Noise            86.07% (1378/1601)
+Language Instructions   87.57% (1346/1537)
+Background Textures     92.01% (990/1076)
+Light Conditions        94.75% (1082/1142)
+
+=== By difficulty ===
+5      54.39% (1133/2083)
+4      73.38% (1384/1886)
+3      80.37% (1683/2094)
+2      87.78% (1933/2202)
+1      92.21% (1516/1644)
+None   98.35% (119/121)
+
+=== Leaderboard row ===
+Model	Camera	Robot	Language	Light	Background	Noise	Layout	Total
+libero_all_stage2_policy_h8_40ep_seed42_from_idm0125_s1v_s2v_noema_epoch_40_pytorch_model.pt	70.73	36.90	87.57	94.75	92.01	86.07	83.21	77.45