Add files using upload-large-folder tool

README.md

@@ -14,8 +14,12 @@ Reactive Diffusion Policy trained on `real_single_flexiv_tactar_flipbox`.
 
 - `checkpoints/at/latest.ckpt`: final Action Tokenizer checkpoint (epoch 600).
 - `checkpoints/ldp/latest.ckpt`: final Latent Diffusion Policy checkpoint (epoch 400, global step 21252).
-- `config/task/*.yaml`: task shape and action-index configuration used for training.
+- `normalization/normalizer.pkl`: exact LDP normalizer computed during training, including action, pose, gripper, wrench, wrist-image and latent-action fields.
+- `config/resolved/{at,ldp}/config.yaml`: fully resolved Hydra configurations saved by each training stage.
+- `config/task/*.yaml`: source task shape and action-index configuration.
+- `config/deployment_config.yaml`: portable relative checkpoint paths and endpoint input contract.
 
 The observation camera is the single wrist view `wrist_img1`. The policy action has 10 dimensions, selected from the converted 20-dimensional VRR action with indices `[0,1,2,3,4,5,6,7,8,19]`.
 
 Both AT and LDP checkpoints are required by the RDP runtime.
+When using the resolved LDP Hydra config after download, override its original absolute `at_load_dir` with `checkpoints/at/latest.ckpt` and set the LDP checkpoint path to `checkpoints/ldp/latest.ckpt`.

config/deployment_config.yaml

@@ -0,0 +1,22 @@
+model: rdp
+task: real_single_flexiv_tactar_flipbox
+camera_keys:
+  - wrist_img1
+checkpoint_paths:
+  at: checkpoints/at/latest.ckpt
+  ldp: checkpoints/ldp/latest.ckpt
+normalizer_path: normalization/normalizer.pkl
+runtime:
+  num_inference_steps: 8
+  horizon: 16
+  n_obs_steps: 2
+  observation_temporal_downsample_ratio: 2
+action:
+  shape: 10
+  source_vrr_indices: [0, 1, 2, 3, 4, 5, 6, 7, 8, 19]
+observation_shapes:
+  wrist_img1: [3, 480, 640]
+  left_robot_tcp_pose: [9]
+  left_robot_gripper_width: [1]
+  left_robot_tcp_wrench: [6]
+

config/resolved/at/config.yaml

@@ -0,0 +1,175 @@
+task:
+  name: real_tactar_flipbox_wrist_wrench_at_10fps
+  dataset_path: /inspire/qb-ilm2/project/robot-body/xuzijun-253108540220/real_single_flexiv_tactar_flipbox_vrr_zarr
+  shape_meta:
+    obs:
+      left_robot_tcp_pose:
+        shape:
+        - 9
+        type: low_dim
+      left_robot_gripper_width:
+        shape:
+        - 1
+        type: low_dim
+      left_robot_tcp_wrench:
+        shape:
+        - 6
+        type: low_dim
+    extended_obs:
+      left_robot_tcp_wrench:
+        shape:
+        - 6
+        type: low_dim
+    action:
+      shape:
+      - 10
+  dataset:
+    _target_: reactive_diffusion_policy.dataset.real_image_tactile_dataset.RealImageTactileDataset
+    shape_meta:
+      obs:
+        left_robot_tcp_pose:
+          shape:
+          - 9
+          type: low_dim
+        left_robot_gripper_width:
+          shape:
+          - 1
+          type: low_dim
+        left_robot_tcp_wrench:
+          shape:
+          - 6
+          type: low_dim
+      extended_obs:
+        left_robot_tcp_wrench:
+          shape:
+          - 6
+          type: low_dim
+      action:
+        shape:
+        - 10
+    dataset_path: ${task.dataset_path}
+    horizon: ${horizon}
+    pad_before: ${eval:'${dataset_obs_steps}-1+${n_latency_steps}'}
+    pad_after: ${eval:'${n_action_steps}-1'}
+    n_obs_steps: ${dataset_obs_steps}
+    obs_temporal_downsample_ratio: ${dataset_obs_temporal_downsample_ratio}
+    n_latency_steps: ${n_latency_steps}
+    seed: 42
+    val_ratio: 0.05
+    max_train_episodes: null
+    delta_action: false
+    relative_action: true
+    transform_params: null
+    action_indices:
+    - 0
+    - 1
+    - 2
+    - 3
+    - 4
+    - 5
+    - 6
+    - 7
+    - 8
+    - 19
+at:
+  horizon: 16
+  n_obs_steps: 2
+  dataset_obs_temporal_downsample_ratio: 1
+  policy:
+    horizon: ${horizon}
+    n_latent_dims: 8
+    mlp_layer_num: 1
+    use_conv_encoder: true
+    conv_latent_dims: 32
+    conv_layer_num: 1
+    use_rnn_decoder: true
+    rnn_latent_dims: 32
+    rnn_layer_num: 1
+    kl_multiplier: 1.0e-06
+    n_embed: 16
+    use_vq: false
+    vqvae_groups: 4
+    eval: false
+    act_scale: 1
+name: train_vae
+_target_: reactive_diffusion_policy.workspace.train_at_workspace.TrainATWorkspace
+task_name: ${task.name}
+shape_meta: ${task.shape_meta}
+exp_name: default
+horizon: ${at.horizon}
+n_obs_steps: ${at.n_obs_steps}
+n_latency_steps: 0
+dataset_obs_steps: ${eval:'${n_obs_steps}*${at.dataset_obs_temporal_downsample_ratio}'}
+dataset_obs_temporal_downsample_ratio: ${at.dataset_obs_temporal_downsample_ratio}
+n_action_steps: ${eval:'${horizon}-${dataset_obs_steps}+1'}
+policy:
+  _target_: reactive_diffusion_policy.model.vae.model.VAE
+  shape_meta: ${task.shape_meta}
+  horizon: ${horizon}
+  n_latent_dims: ${at.policy.n_latent_dims}
+  mlp_layer_num: ${at.policy.mlp_layer_num}
+  use_conv_encoder: ${at.policy.use_conv_encoder}
+  conv_latent_dims: ${at.policy.conv_latent_dims}
+  conv_layer_num: ${at.policy.conv_layer_num}
+  use_rnn_decoder: ${at.policy.use_rnn_decoder}
+  rnn_latent_dims: ${at.policy.rnn_latent_dims}
+  rnn_layer_num: ${at.policy.rnn_layer_num}
+  kl_multiplier: ${at.policy.kl_multiplier}
+  n_embed: ${at.policy.n_embed}
+  use_vq: ${at.policy.use_vq}
+  vqvae_groups: ${at.policy.vqvae_groups}
+  eval: ${at.policy.eval}
+  act_scale: ${at.policy.act_scale}
+dataloader:
+  batch_size: 64
+  num_workers: 4
+  shuffle: true
+  pin_memory: true
+  persistent_workers: true
+val_dataloader:
+  batch_size: 64
+  num_workers: 4
+  shuffle: false
+  pin_memory: true
+  persistent_workers: true
+optimizer:
+  _target_: torch.optim.AdamW
+  lr: 0.001
+  weight_decay: 0.0001
+training:
+  device: cuda:0
+  seed: 42
+  debug: false
+  resume: true
+  lr_scheduler: cosine
+  lr_warmup_steps: 100
+  num_epochs: 601
+  gradient_accumulate_every: 1
+  use_ema: false
+  freeze_encoder: false
+  rollout_every: -1
+  checkpoint_every: 10
+  val_every: 1
+  sample_every: -1
+  max_train_steps: null
+  max_val_steps: null
+  tqdm_interval_sec: 1.0
+logging:
+  project: at
+  resume: true
+  mode: offline
+  name: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
+  tags:
+  - ${name}
+  - ${task_name}
+  - ${exp_name}
+  id: null
+  group: null
+checkpoint:
+  topk:
+    monitor_key: train_loss
+    mode: min
+    k: 1
+    format_str: epoch={epoch:04d}-train_loss={train_loss:.6f}.ckpt
+  save_last_ckpt: true
+  save_last_snapshot: false

config/resolved/at/overrides.yaml

@@ -0,0 +1,4 @@
+- task=real_tactar_flipbox_wrist_wrench_at_10fps
+- at=at_tactar_10fps
+- task.dataset_path=/inspire/qb-ilm2/project/robot-body/xuzijun-253108540220/real_single_flexiv_tactar_flipbox_vrr_zarr
+- logging.mode=offline

config/resolved/ldp/config.yaml

@@ -0,0 +1,270 @@
+task:
+  name: real_tactar_flipbox_wrist_wrench_ldp_10fps
+  dataset_path: /inspire/qb-ilm2/project/robot-body/xuzijun-253108540220/real_single_flexiv_tactar_flipbox_vrr_zarr
+  image_shape:
+  - 3
+  - 480
+  - 640
+  shape_meta:
+    obs:
+      wrist_img1:
+        shape: ${task.image_shape}
+        type: rgb
+      left_robot_tcp_pose:
+        shape:
+        - 9
+        type: low_dim
+      left_robot_gripper_width:
+        shape:
+        - 1
+        type: low_dim
+      left_robot_tcp_wrench:
+        shape:
+        - 6
+        type: low_dim
+    extended_obs:
+      left_robot_tcp_wrench:
+        shape:
+        - 6
+        type: low_dim
+    action:
+      shape:
+      - 10
+  dataset:
+    _target_: reactive_diffusion_policy.dataset.real_image_tactile_latent_diffusion_dataset.RealImageTactileLatentDiffusionDataset
+    shape_meta:
+      obs:
+        wrist_img1:
+          shape: ${task.image_shape}
+          type: rgb
+        left_robot_tcp_pose:
+          shape:
+          - 9
+          type: low_dim
+        left_robot_gripper_width:
+          shape:
+          - 1
+          type: low_dim
+        left_robot_tcp_wrench:
+          shape:
+          - 6
+          type: low_dim
+      extended_obs:
+        left_robot_tcp_wrench:
+          shape:
+          - 6
+          type: low_dim
+      action:
+        shape:
+        - 10
+    dataset_path: ${task.dataset_path}
+    horizon: ${horizon}
+    pad_before: ${eval:'${dataset_obs_steps}-1+${n_latency_steps}'}
+    pad_after: ${eval:'${n_action_steps}-1'}
+    n_obs_steps: ${dataset_obs_steps}
+    obs_temporal_downsample_ratio: ${dataset_obs_temporal_downsample_ratio}
+    n_latency_steps: ${n_latency_steps}
+    seed: 42
+    val_ratio: 0.05
+    max_train_episodes: null
+    delta_action: false
+    relative_action: true
+    transform_params: null
+    action_indices:
+    - 0
+    - 1
+    - 2
+    - 3
+    - 4
+    - 5
+    - 6
+    - 7
+    - 8
+    - 19
+    use_latent_action_before_vq: false
+    at:
+      _target_: reactive_diffusion_policy.model.vae.model.VAE
+      shape_meta: ${shape_meta}
+      horizon: ${at.policy.horizon}
+      n_latent_dims: ${at.policy.n_latent_dims}
+      mlp_layer_num: ${at.policy.mlp_layer_num}
+      use_conv_encoder: ${at.policy.use_conv_encoder}
+      conv_latent_dims: ${at.policy.conv_latent_dims}
+      conv_layer_num: ${at.policy.conv_layer_num}
+      use_rnn_decoder: ${at.policy.use_rnn_decoder}
+      rnn_latent_dims: ${at.policy.rnn_latent_dims}
+      rnn_layer_num: ${at.policy.rnn_layer_num}
+      use_vq: ${at.policy.use_vq}
+      n_embed: ${at.policy.n_embed}
+      vqvae_groups: ${at.policy.vqvae_groups}
+      eval: true
+      load_dir: ${at_load_dir}
+      act_scale: ${at.policy.act_scale}
+at:
+  horizon: 16
+  n_obs_steps: 2
+  dataset_obs_temporal_downsample_ratio: 1
+  policy:
+    horizon: ${horizon}
+    n_latent_dims: 8
+    mlp_layer_num: 1
+    use_conv_encoder: true
+    conv_latent_dims: 32
+    conv_layer_num: 1
+    use_rnn_decoder: true
+    rnn_latent_dims: 32
+    rnn_layer_num: 1
+    kl_multiplier: 1.0e-06
+    n_embed: 16
+    use_vq: false
+    vqvae_groups: 4
+    eval: false
+    act_scale: 1
+name: train_latent_diffusion_unet_image
+_target_: reactive_diffusion_policy.workspace.train_diffusion_unet_image_workspace.TrainDiffusionUnetImageWorkspace
+task_name: ${task.name}
+shape_meta: ${task.shape_meta}
+exp_name: default
+horizon: ${at.horizon}
+n_obs_steps: ${at.n_obs_steps}
+n_latency_steps: 0
+obs_temporal_downsample_ratio: 2
+dataset_obs_steps: ${eval:'${at.n_obs_steps}*${at.dataset_obs_temporal_downsample_ratio}'}
+dataset_obs_temporal_downsample_ratio: ${at.dataset_obs_temporal_downsample_ratio}
+n_action_steps: ${eval:'${horizon}-${dataset_obs_steps}+1'}
+past_action_visible: false
+keypoint_visible_rate: 1.0
+at_load_dir: /inspire/qb-ilm2/project/robot-body/xuzijun-253108540220/reactive_diffusion_policy/runs/tactar_flipbox_rdp/20260801_0458/at/checkpoints/latest.ckpt
+use_latent_action_before_vq: false
+obs_as_global_cond: true
+policy:
+  _target_: reactive_diffusion_policy.policy.latent_diffusion_unet_image_policy.LatentDiffusionUnetImagePolicy
+  shape_meta: ${shape_meta}
+  noise_scheduler:
+    _target_: diffusers.schedulers.scheduling_ddim.DDIMScheduler
+    num_train_timesteps: 100
+    beta_start: 0.0001
+    beta_end: 0.02
+    beta_schedule: squaredcos_cap_v2
+    clip_sample: true
+    set_alpha_to_one: true
+    steps_offset: 0
+    prediction_type: epsilon
+  obs_encoder:
+    _target_: reactive_diffusion_policy.model.vision.multi_image_obs_encoder.MultiImageObsEncoder
+    shape_meta: ${shape_meta}
+    rgb_model:
+      _target_: reactive_diffusion_policy.model.vision.model_getter.get_resnet
+      name: resnet18
+      weights: null
+    resize_shape:
+    - 240
+    - 320
+    random_transforms:
+    - type: RandomCrop
+      ratio: 0.9
+    use_group_norm: true
+    share_rgb_model: false
+    imagenet_norm: true
+  use_latent_action_before_vq: ${use_latent_action_before_vq}
+  at:
+    _target_: reactive_diffusion_policy.model.vae.model.VAE
+    shape_meta: ${shape_meta}
+    horizon: ${at.policy.horizon}
+    n_latent_dims: ${at.policy.n_latent_dims}
+    mlp_layer_num: ${at.policy.mlp_layer_num}
+    use_conv_encoder: ${at.policy.use_conv_encoder}
+    conv_latent_dims: ${at.policy.conv_latent_dims}
+    conv_layer_num: ${at.policy.conv_layer_num}
+    use_rnn_decoder: ${at.policy.use_rnn_decoder}
+    rnn_latent_dims: ${at.policy.rnn_latent_dims}
+    rnn_layer_num: ${at.policy.rnn_layer_num}
+    use_vq: ${at.policy.use_vq}
+    n_embed: ${at.policy.n_embed}
+    vqvae_groups: ${at.policy.vqvae_groups}
+    eval: true
+    load_dir: ${at_load_dir}
+    act_scale: ${at.policy.act_scale}
+  horizon: ${horizon}
+  n_action_steps: ${eval:'${n_action_steps}+${n_latency_steps}'}
+  n_obs_steps: ${n_obs_steps}
+  num_inference_steps: 100
+  obs_as_global_cond: ${obs_as_global_cond}
+  diffusion_step_embed_dim: 128
+  down_dims:
+  - 512
+  - 1024
+  - 2048
+  kernel_size: 5
+  n_groups: 8
+  cond_predict_scale: true
+ema:
+  _target_: reactive_diffusion_policy.model.diffusion.ema_model.EMAModel
+  update_after_step: 0
+  inv_gamma: 1.0
+  power: 0.75
+  min_value: 0.0
+  max_value: 0.9999
+dataloader:
+  batch_size: 64
+  num_workers: 0
+  shuffle: true
+  pin_memory: true
+  persistent_workers: false
+val_dataloader:
+  batch_size: 64
+  num_workers: 0
+  shuffle: false
+  pin_memory: true
+  persistent_workers: false
+optimizer:
+  _target_: torch.optim.AdamW
+  lr: 0.0001
+  betas:
+  - 0.95
+  - 0.999
+  eps: 1.0e-08
+  weight_decay: 1.0e-06
+  encoder_weight_decay: 1.0e-06
+training:
+  device: cuda:0
+  seed: 42
+  debug: false
+  resume: true
+  lr_scheduler: cosine
+  lr_warmup_steps: 500
+  num_epochs: 401
+  gradient_accumulate_every: 1
+  use_ema: true
+  freeze_encoder: false
+  rollout_every: 50
+  checkpoint_every: 10
+  val_every: 1
+  sample_every: 10
+  max_train_steps: null
+  max_val_steps: null
+  tqdm_interval_sec: 1.0
+  layer_decay: 1.0
+  encoder_lr_coefficient: 0.1
+logging:
+  project: latent_diffusion_policy_tactile
+  resume: true
+  mode: offline
+  name: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}
+  tags:
+  - ${name}
+  - ${task_name}
+  - ${exp_name}
+  id: null
+  group: null
+checkpoint:
+  topk:
+    monitor_key: train_loss
+    mode: min
+    k: 1
+    format_str: epoch={epoch:04d}-train_loss={train_loss:.3f}.ckpt
+  save_last_ckpt: true
+  save_last_snapshot: false
+multi_run:
+  run_dir: data/outputs/${now:%Y.%m.%d}/${now:%H.%M.%S}_${name}_${task_name}
+  wandb_name_base: ${now:%Y.%m.%d-%H.%M.%S}_${name}_${task_name}

config/resolved/ldp/overrides.yaml

@@ -0,0 +1,9 @@
+- task=real_tactar_flipbox_wrist_wrench_ldp_10fps
+- at=at_tactar_10fps
+- at_load_dir=/inspire/qb-ilm2/project/robot-body/xuzijun-253108540220/reactive_diffusion_policy/runs/tactar_flipbox_rdp/20260801_0458/at/checkpoints/latest.ckpt
+- task.dataset_path=/inspire/qb-ilm2/project/robot-body/xuzijun-253108540220/real_single_flexiv_tactar_flipbox_vrr_zarr
+- dataloader.num_workers=0
+- dataloader.persistent_workers=False
+- val_dataloader.num_workers=0
+- val_dataloader.persistent_workers=False
+- logging.mode=offline

normalization/normalizer.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5552adfe78956d6973aefc76121ded8bcc6b0447a02d8b28d1b3fcf7c93073b1
+size 14383