feat(train): 跑通训练脚本

roboimi/demos/vla_scripts/train_vla.py

@@ -7,115 +7,27 @@ from tqdm import tqdm
 from omegaconf import DictConfig, OmegaConf
 from torch.utils.data import DataLoader
 from torch.optim import AdamW
+from pathlib import Path
 
-# 确保导入路径正确
+# Ensure correct import path
 sys.path.append(os.getcwd())
 
-from roboimi.vla.agent import VLAAgent
 from hydra.utils import instantiate
 
 log = logging.getLogger(__name__)
 
-@hydra.main(version_base=None, config_path="../../vla/conf", config_name="config")
-def main(cfg: DictConfig):
-    print(OmegaConf.to_yaml(cfg))
-    log.info(f"🚀 Starting VLA Training with Real Data (Device: {cfg.train.device})")
-
-    # --- 1. 实例化 Dataset & DataLoader ---
-    # Hydra 根据 conf/data/custom_hdf5.yaml 实例化类
-    dataset = instantiate(cfg.data)
-    
-    dataloader = DataLoader(
-        dataset,
-        batch_size=cfg.train.batch_size,
-        shuffle=True,
-        num_workers=cfg.train.num_workers,
-        pin_memory=(cfg.train.device != "cpu")
-    )
-    log.info(f"✅ Dataset loaded. Size: {len(dataset)}")
-
-    # --- 2. 实例化 Agent ---
-    agent: VLAAgent = instantiate(cfg.agent)
-    agent.to(cfg.train.device)
-    agent.train()
-
-    optimizer = AdamW(agent.parameters(), lr=cfg.train.lr)
-
-    # --- 3. Training Loop ---
-    # 使用一个无限迭代器或者 epoch 循环
-    data_iter = iter(dataloader)
-    pbar = tqdm(range(cfg.train.max_steps), desc="Training")
-    
-    for step in pbar:
-        try:
-            batch = next(data_iter)
-        except StopIteration:
-            #而在 epoch 结束时重新开始
-            data_iter = iter(dataloader)
-            batch = next(data_iter)
-
-        # Move to device
-        # 注意：这里需要递归地将字典里的 tensor 移到 GPU
-        batch = recursive_to_device(batch, cfg.train.device)
-
-        # --- 4. Adapter Layer (适配层) ---
-        # Dataset 返回的是具体的相机 key (如 'agentview_image' 或 'top')
-        # Agent 期望的是通用的 'image'
-        # 我们在这里做一个映射，模拟多模态融合前的处理
-        
-        # 假设我们只用配置里的第一个 key 作为主视觉
-        # primary_cam_key = cfg.data.obs_keys[0] 
-        
-        # Dataset 返回 shape: (B, Obs_Horizon, C, H, W)
-        # DebugBackbone 期望: (B, C, H, W) 或者 (B, Seq, Dim)
-        # 这里我们取 Obs_Horizon 的最后一帧 (Current Frame)
-        # input_img = batch['obs'][primary_cam_key][:, -1, :, :, :]
-        
-        # agent_input = {
-        #     "obs": {
-        #         "image": input_img,
-        #         "text": batch["language"] # 传递语言指令
-        #     },
-        #     "actions": batch["actions"] # (B, Chunk, Dim)
-        # }
-        agent_input = {
-            "action": batch["action"],
-            "qpos": batch["qpos"],
-            "images": {}
-        }
-
-        for cam_name in cfg.data.camera_names:
-            key = f"image_{cam_name}"
-            agent_input["images"][cam_name] = batch[key].squeeze(1)
-
-        # --- 5. Forward & Backward ---
-        outputs = agent(agent_input)
-        
-        # 处理 Loss 掩码 (如果在真实训练中，需要在这里应用 action_mask)
-        # 目前 DebugHead 内部直接算了 MSE，还没用 mask，我们在下一阶段优化 Policy 时加上
-        loss = outputs['loss']
-
-        optimizer.zero_grad()
-        loss.backward()
-        optimizer.step()
-
-        if step % cfg.train.log_freq == 0:
-            pbar.set_postfix({"loss": f"{loss.item():.4f}"})
-
-    log.info("✅ Training Loop with Real HDF5 Finished!")
-
-# --- 6. Save Checkpoint ---
-    save_dir = "checkpoints"
-    os.makedirs(save_dir, exist_ok=True)
-    save_path = os.path.join(save_dir, "vla_model_final.pt")
-    
-    # 保存整个 Agent 的 state_dict
-    torch.save(agent.state_dict(), save_path)
-    log.info(f"💾 Model saved to {save_path}")
-
-    log.info("✅ Training Loop Finished!")
 
 def recursive_to_device(data, device):
+    """
+    Recursively move nested dictionaries/lists of tensors to specified device.
+
+    Args:
+        data: Dictionary, list, or tensor
+        device: Target device (e.g., 'cuda', 'cpu')
+
+    Returns:
+        Data structure with all tensors moved to device
+    """
     if isinstance(data, torch.Tensor):
         return data.to(device)
     elif isinstance(data, dict):
@@ -124,5 +36,193 @@ def recursive_to_device(data, device):
         return [recursive_to_device(v, device) for v in data]
     return data
 
+
+@hydra.main(version_base=None, config_path="../../vla/conf", config_name="config")
+def main(cfg: DictConfig):
+    """
+    VLA Training Script with ResNet Backbone and Diffusion Policy.
+
+    This script:
+    1. Loads dataset from HDF5 files
+    2. Instantiates VLAAgent with ResNet vision encoder
+    3. Trains diffusion-based action prediction
+    4. Saves checkpoints periodically
+    """
+
+    # Print configuration
+    print("=" * 80)
+    print("VLA Training Configuration:")
+    print("=" * 80)
+    print(OmegaConf.to_yaml(cfg))
+    print("=" * 80)
+
+    log.info(f"🚀 Starting VLA Training (Device: {cfg.train.device})")
+
+    # Create checkpoint directory
+    checkpoint_dir = Path("checkpoints")
+    checkpoint_dir.mkdir(exist_ok=True)
+
+    # =========================================================================
+    # 1. Instantiate Dataset & DataLoader
+    # =========================================================================
+    log.info("📦 Loading dataset...")
+    try:
+        dataset = instantiate(cfg.data)
+        log.info(f"✅ Dataset loaded successfully. Total samples: {len(dataset)}")
+    except Exception as e:
+        log.error(f"❌ Failed to load dataset: {e}")
+        raise
+
+    dataloader = DataLoader(
+        dataset,
+        batch_size=cfg.train.batch_size,
+        shuffle=True,
+        num_workers=cfg.train.num_workers,
+        pin_memory=(cfg.train.device != "cpu"),
+        drop_last=True  # Drop incomplete batches for stable training
+    )
+    log.info(f"✅ DataLoader created. Batches per epoch: {len(dataloader)}")
+
+    # =========================================================================
+    # 2. Instantiate VLA Agent
+    # =========================================================================
+    log.info("🤖 Initializing VLA Agent...")
+    try:
+        agent = instantiate(cfg.agent)
+        agent.to(cfg.train.device)
+        agent.train()
+        log.info(f"✅ Agent initialized and moved to {cfg.train.device}")
+
+        # Count parameters
+        total_params = sum(p.numel() for p in agent.parameters())
+        trainable_params = sum(p.numel() for p in agent.parameters() if p.requires_grad)
+        log.info(f"📊 Total parameters: {total_params:,}")
+        log.info(f"📊 Trainable parameters: {trainable_params:,}")
+
+    except Exception as e:
+        log.error(f"❌ Failed to initialize agent: {e}")
+        raise
+
+    # =========================================================================
+    # 3. Setup Optimizer
+    # =========================================================================
+    optimizer = AdamW(agent.parameters(), lr=cfg.train.lr, weight_decay=1e-5)
+    log.info(f"🔧 Optimizer: AdamW (lr={cfg.train.lr})")
+
+    # =========================================================================
+    # 4. Training Loop
+    # =========================================================================
+    log.info("🏋️ Starting training loop...")
+
+    data_iter = iter(dataloader)
+    pbar = tqdm(range(cfg.train.max_steps), desc="Training", ncols=100)
+
+    best_loss = float('inf')
+
+    for step in pbar:
+        try:
+            batch = next(data_iter)
+        except StopIteration:
+            # Restart iterator when epoch ends
+            data_iter = iter(dataloader)
+            batch = next(data_iter)
+
+        # =====================================================================
+        # Move batch to device
+        # =====================================================================
+        batch = recursive_to_device(batch, cfg.train.device)
+
+        # =====================================================================
+        # Prepare agent input
+        # =====================================================================
+        # Dataset returns: {action, qpos, image_<cam_name>, ...}
+        # Agent expects: {images: dict, qpos: tensor, action: tensor}
+
+        # Extract images into a dictionary
+        images = {}
+        for cam_name in cfg.data.camera_names:
+            key = f"image_{cam_name}"
+            if key in batch:
+                images[cam_name] = batch[key]  # (B, obs_horizon, C, H, W)
+
+        # Prepare agent input
+        agent_input = {
+            'images': images,                      # Dict of camera images
+            'qpos': batch['qpos'],            # (B, obs_horizon, obs_dim)
+            'action': batch['action']             # (B, pred_horizon, action_dim)
+        }
+
+        # =====================================================================
+        # Forward pass & compute loss
+        # =====================================================================
+        try:
+            loss = agent.compute_loss(agent_input)
+        except Exception as e:
+            log.error(f"❌ Forward pass failed at step {step}: {e}")
+            raise
+
+        # =====================================================================
+        # Backward pass & optimization
+        # =====================================================================
+        optimizer.zero_grad()
+        loss.backward()
+
+        # Gradient clipping for stable training
+        torch.nn.utils.clip_grad_norm_(agent.parameters(), max_norm=1.0)
+
+        optimizer.step()
+
+        # =====================================================================
+        # Logging
+        # =====================================================================
+        if step % cfg.train.log_freq == 0:
+            pbar.set_postfix({
+                "loss": f"{loss.item():.4f}",
+                "best_loss": f"{best_loss:.4f}"
+            })
+            log.info(f"Step {step}/{cfg.train.max_steps} | Loss: {loss.item():.4f}")
+
+        # =====================================================================
+        # Checkpoint saving
+        # =====================================================================
+        if step > 0 and step % cfg.train.save_freq == 0:
+            checkpoint_path = checkpoint_dir / f"vla_model_step_{step}.pt"
+            torch.save({
+                'step': step,
+                'model_state_dict': agent.state_dict(),
+                'optimizer_state_dict': optimizer.state_dict(),
+                'loss': loss.item(),
+            }, checkpoint_path)
+            log.info(f"💾 Checkpoint saved: {checkpoint_path}")
+
+            # Save best model
+            if loss.item() < best_loss:
+                best_loss = loss.item()
+                best_model_path = checkpoint_dir / "vla_model_best.pt"
+                torch.save({
+                    'step': step,
+                    'model_state_dict': agent.state_dict(),
+                    'optimizer_state_dict': optimizer.state_dict(),
+                    'loss': loss.item(),
+                }, best_model_path)
+                log.info(f"🌟 Best model updated: {best_model_path} (loss: {best_loss:.4f})")
+
+    # =========================================================================
+    # 5. Save Final Model
+    # =========================================================================
+    final_model_path = checkpoint_dir / "vla_model_final.pt"
+    torch.save({
+        'step': cfg.train.max_steps,
+        'model_state_dict': agent.state_dict(),
+        'optimizer_state_dict': optimizer.state_dict(),
+        'loss': loss.item(),
+    }, final_model_path)
+    log.info(f"💾 Final model saved: {final_model_path}")
+
+    log.info("✅ Training completed successfully!")
+    log.info(f"📊 Final Loss: {loss.item():.4f}")
+    log.info(f"📊 Best Loss: {best_loss:.4f}")
+
+
 if __name__ == "__main__":
-    main()
+    main()

roboimi/vla/RESNET_TRAINING_GUIDE.md

@@ -0,0 +1,238 @@
+# ResNet VLA Training Guide
+
+This guide explains how to train the VLA agent with ResNet backbone and action_dim=16, obs_dim=16.
+
+## Configuration Overview
+
+### 1. Backbone Configuration
+**File**: `roboimi/vla/conf/backbone/resnet.yaml`
+- Model: microsoft/resnet-18
+- Output dim: 1024 (512 channels × 2 from SpatialSoftmax)
+- Frozen by default for faster training
+
+### 2. Agent Configuration
+**File**: `roboimi/vla/conf/agent/resnet_diffusion.yaml`
+- Vision backbone: ResNet-18 with SpatialSoftmax
+- Action dimension: 16
+- Observation dimension: 16
+- Prediction horizon: 16 steps
+- Observation horizon: 2 steps
+- Diffusion steps: 100
+- Number of cameras: 2
+
+### 3. Dataset Configuration
+**File**: `roboimi/vla/conf/data/resnet_dataset.yaml`
+- Dataset class: RobotDiffusionDataset
+- Prediction horizon: 16
+- Observation horizon: 2
+- Camera names: [r_vis, top]
+- Normalization: gaussian (mean/std)
+
+### 4. Training Configuration
+**File**: `roboimi/vla/conf/config.yaml`
+- Batch size: 8
+- Learning rate: 1e-4
+- Max steps: 10000
+- Log frequency: 100 steps
+- Save frequency: 1000 steps
+- Device: cuda
+- Num workers: 4
+
+## Prerequisites
+
+### 1. Prepare Dataset
+Your dataset should be organized as:
+```
+/path/to/your/dataset/
+├── episode_0.hdf5
+├── episode_1.hdf5
+├── ...
+└── data_stats.pkl
+```
+
+Each HDF5 file should contain:
+```
+episode_N.hdf5
+├── action              # (T, 16) float32
+└── observations/
+    ├── qpos           # (T, 16) float32
+    └── images/
+        ├── r_vis/     # (T, H, W, 3) uint8
+        └── top/       # (T, H, W, 3) uint8
+```
+
+### 2. Generate Dataset Statistics
+Create `data_stats.pkl` with:
+```python
+import pickle
+import numpy as np
+
+stats = {
+    'action': {
+        'mean': np.zeros(16),
+        'std': np.ones(16)
+    },
+    'qpos': {
+        'mean': np.zeros(16),
+        'std': np.ones(16)
+    }
+}
+
+with open('/path/to/your/dataset/data_stats.pkl', 'wb') as f:
+    pickle.dump(stats, f)
+```
+
+Or use the provided script:
+```bash
+python -m roboimi.vla.scripts.calculate_stats --dataset_dir /path/to/your/dataset
+```
+
+## Usage
+
+### 1. Update Dataset Path
+Edit `roboimi/vla/conf/data/resnet_dataset.yaml`:
+```yaml
+dataset_dir: "/path/to/your/dataset"  # CHANGE THIS
+camera_names:
+  - r_vis  # CHANGE TO YOUR CAMERA NAMES
+  - top
+```
+
+### 2. Run Training
+```bash
+# Basic training
+python roboimi/demos/vla_scripts/train_vla.py
+
+# Override configurations
+python roboimi/demos/vla_scripts/train_vla.py train.batch_size=16
+python roboimi/demos/vla_scripts/train_vla.py train.device=cpu
+python roboimi/demos/vla_scripts/train_vla.py train.max_steps=20000
+python roboimi/demos/vla_scripts/train_vla.py data.dataset_dir=/custom/path
+
+# Debug mode (CPU, small batch, few steps)
+python roboimi/demos/vla_scripts/train_vla.py \
+    train.device=cpu \
+    train.batch_size=2 \
+    train.max_steps=10 \
+    train.num_workers=0
+```
+
+### 3. Monitor Training
+Checkpoints are saved to:
+- `checkpoints/vla_model_step_1000.pt` - Periodic checkpoints
+- `checkpoints/vla_model_best.pt` - Best model (lowest loss)
+- `checkpoints/vla_model_final.pt` - Final model
+
+## Architecture Details
+
+### Data Flow
+1. **Input**: Images from multiple cameras + proprioception (qpos)
+2. **Vision Encoder**: ResNet-18 → SpatialSoftmax → (B, T, 1024) per camera
+3. **Feature Concatenation**: All cameras + qpos → Global conditioning
+4. **Diffusion Policy**: 1D U-Net predicts noise on action sequences
+5. **Output**: Clean action sequence (B, 16, 16)
+
+### Training Process
+1. Sample random timestep t from [0, 100]
+2. Add noise to ground truth actions
+3. Predict noise using vision + proprioception conditioning
+4. Compute MSE loss between predicted and actual noise
+5. Backpropagate and update weights
+
+### Inference Process
+1. Extract visual features from current observation
+2. Start with random noise action sequence
+3. Iteratively denoise over 10 steps (DDPM scheduler)
+4. Return clean action sequence
+
+## Common Issues
+
+### Issue: Out of Memory
+**Solution**: Reduce batch size or use CPU
+```bash
+python train_vla.py train.batch_size=4 train.device=cpu
+```
+
+### Issue: Dataset not found
+**Solution**: Check dataset_dir path in config
+```bash
+python train_vla.py data.dataset_dir=/absolute/path/to/dataset
+```
+
+### Issue: Camera names mismatch
+**Solution**: Update camera_names in data config
+```yaml
+# roboimi/vla/conf/data/resnet_dataset.yaml
+camera_names:
+  - your_camera_1
+  - your_camera_2
+```
+
+### Issue: data_stats.pkl missing
+**Solution**: Generate statistics file
+```bash
+python -m roboimi.vla.scripts.calculate_stats --dataset_dir /path/to/dataset
+```
+
+## Model Files Created
+
+```
+roboimi/vla/
+├── conf/
+│   ├── config.yaml (UPDATED)
+│   ├── backbone/
+│   │   └── resnet.yaml (NEW)
+│   ├── agent/
+│   │   └── resnet_diffusion.yaml (NEW)
+│   └── data/
+│       └── resnet_dataset.yaml (NEW)
+├── models/
+│   └── backbones/
+│       ├── __init__.py (UPDATED - added resnet export)
+│       └── resnet.py (EXISTING)
+└── demos/vla_scripts/
+    └── train_vla.py (REWRITTEN)
+```
+
+## Next Steps
+
+1. **Prepare your dataset** in the required HDF5 format
+2. **Update dataset_dir** in `roboimi/vla/conf/data/resnet_dataset.yaml`
+3. **Run training** with `python roboimi/demos/vla_scripts/train_vla.py`
+4. **Monitor checkpoints** in `checkpoints/` directory
+5. **Evaluate** the trained model using the best checkpoint
+
+## Advanced Configuration
+
+### Use Different ResNet Variant
+Edit `roboimi/vla/conf/agent/resnet_diffusion.yaml`:
+```yaml
+vision_backbone:
+  model_name: "microsoft/resnet-50"  # or resnet-34, resnet-101
+```
+
+### Adjust Diffusion Steps
+```yaml
+# More steps = better quality, slower training
+diffusion_steps: 200  # default: 100
+```
+
+### Change Horizons
+```yaml
+pred_horizon: 32  # Predict more future steps
+obs_horizon: 4    # Use more history
+```
+
+### Multi-GPU Training
+```bash
+# Use CUDA device 1
+python train_vla.py train.device=cuda:1
+
+# For multi-GPU, use torch.distributed (requires code modification)
+```
+
+## References
+
+- ResNet Paper: https://arxiv.org/abs/1512.03385
+- Diffusion Policy: https://diffusion-policy.cs.columbia.edu/
+- VLA Framework Documentation: See CLAUDE.md in project root

roboimi/vla/agent.py

@@ -2,111 +2,127 @@ import torch
 import torch.nn as nn
 from typing import Dict, Optional, Any
 from roboimi.vla.core.interfaces import VLABackbone, VLAProjector, VLAHead
+from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
+from roboimi.vla.models.heads.diffusion import ConditionalUnet1D
 
 class VLAAgent(nn.Module):
 
     def __init__(
         self,
-        backbone: VLABackbone,
-        projector: VLAProjector,
-        head: VLAHead,
-        state_encoder: nn.Module
+        vision_backbone,      # 你之前定义的 ResNet 类
+        action_dim,           # 机器人动作维度 (例如 7: xyz + rpy + gripper)
+        obs_dim,              # 本体感知维度 (例如 关节角度)
+        pred_horizon=16,      # 预测未来多少步动作
+        obs_horizon=4,        # 使用多少步历史观测
+        diffusion_steps=100,
+        num_cams=2,          # 视觉输入的摄像头数量
     ):
         super().__init__()
-        self.backbone = backbone
-        self.projector = projector
-        self.head = head
-        self.state_encoder = state_encoder
+        self.vision_encoder = vision_backbone
+        single_img_feat_dim = self.vision_encoder.output_dim
+        total_vision_dim = single_img_feat_dim * num_cams * obs_horizon
+        total_prop_dim = obs_dim * obs_horizon
+        self.global_cond_dim = total_vision_dim + total_prop_dim
 
-    def forward(self, batch: Dict[str, Any]) -> Dict[str, torch.Tensor]:
-
-        action = batch["action"]
-        state = batch["qpos"]
-        images = batch["images"]
-
-        state_emb = self.state_encoder(state)
-
-        # 2. Project Features
-        # Shape: (B, Seq, Head_Dim)
-        embeddings = self.projector(features)
-
-        # 3. Compute Action/Loss
-        # We pass actions if they exist (training mode)
-        actions = batch.get('actions', None)
-        outputs = self.head(embeddings=embeddings, actions=actions)
-
-        return outputs
-
-# # roboimi/vla/agent.py
-
-# import torch
-# import torch.nn as nn
-# from typing import Optional, Dict, Union
-
-# class VLAAgent(nn.Module):
-#     def __init__(self, 
-#                  vlm_backbone: nn.Module,
-#                  img_projector: nn.Module,
-#                  action_head: nn.Module,
-#                  state_dim: int,
-#                  embed_dim: int):
-#         super().__init__()
-#         self.vlm_backbone = vlm_backbone
-#         self.img_projector = img_projector
-#         self.action_head = action_head
+        self.noise_scheduler = DDPMScheduler(
+                    num_train_timesteps=diffusion_steps,
+                    beta_schedule='squaredcos_cap_v2', # 机器人任务常用的 schedule
+                    clip_sample=True,
+                    prediction_type='epsilon' # 预测噪声
+                )
         
-#         # 简单的状态编码器 (通常不需要复杂的 config，直接写在这里即可)
-#         self.state_encoder = nn.Sequential(
-#             nn.Linear(state_dim, embed_dim),
-#             nn.Mish(),
-#             nn.Linear(embed_dim, embed_dim)
-#         )
+        self.noise_pred_net = ConditionalUnet1D(
+            input_dim=action_dim,
+            global_cond_dim=self.global_cond_dim
+        )
 
-#     def forward(self, 
-#                 images: torch.Tensor, 
-#                 state: torch.Tensor, 
-#                 text: Optional[Union[str, list]] = None, 
-#                 actions: Optional[torch.Tensor] = None) -> Union[torch.Tensor, Dict]:
-#         """
-#         Args:
-#             images: [Batch, Obs_Horizon, C, H, W]  注意: 这里需要处理时间维度
-#             state:  [Batch, Obs_Horizon, State_Dim]
-#             text:   Optional text instructions
-#             actions: [Batch, Pred_Horizon, Action_Dim] (Training only)
+    # ==========================
+    # 训练阶段 (Training)
+    # ==========================
+    def compute_loss(self, batch):
+        """
+        batch: 包含 images, qpos (proprioception), action
+        """
+        gt_actions = batch['action'] # Shape: (B, Horizon, Action_Dim)
+        B = gt_actions.shape[0]
+        images = batch['images']
+        proprioception = batch['qpos'].view(B, -1) # (B, obs_horizon * obs_dim)
+
+
+        # 1. 提取视觉特征
+        visual_features = self.vision_encoder(images).view(B, -1) # (B, vision_dim)
+        
+        # 2. 融合特征 -> 全局条件 (Global Conditioning)
+        global_cond = torch.cat([visual_features, proprioception], dim=-1)
+
+        # 3. 采样噪声
+        noise = torch.randn_like(gt_actions)
+        
+        # 4. 随机采样时间步 (Timesteps)
+        timesteps = torch.randint(
+            0, self.noise_scheduler.config.num_train_timesteps, 
+            (B,), device=gt_actions.device
+        ).long()
+
+        # 5. 给动作加噪 (Forward Diffusion)
+        noisy_actions = self.noise_scheduler.add_noise(
+            gt_actions, noise, timesteps
+        )
+
+        # 6. 网络预测噪声
+        # 注意：U-Net 1D 通常期望 channel 在中间: (B, C, T)
+        # noisy_actions_inp = noisy_actions.permute(0, 2, 1) 
+        
+        pred_noise = self.noise_pred_net(
+            sample=noisy_actions, 
+            timestep=timesteps, 
+            global_cond=global_cond
+        )
+        
+        # 还原维度 (B, T, C)
+        pred_noise = pred_noise.permute(0, 2, 1)
+
+        # 7. 计算 Loss (MSE)
+        loss = nn.functional.mse_loss(pred_noise, noise)
+        return loss
+
+    # ==========================
+    # 推理阶段 (Inference)
+    # ==========================
+    @torch.no_grad()
+    def predict_action(self, images, proprioception):
+        B = 1 # 假设单次推理
+        
+        # 1. 提取当前观测特征 (只做一次)
+        visual_features = self.vision_encoder(images).view(B, -1)
+        proprioception = proprioception.view(B, -1)
+        global_cond = torch.cat([visual_features, proprioception], dim=-1)
+
+        # 2. 初始化纯高斯噪声动作
+        # Shape: (B, Horizon, Action_Dim)
+        current_actions = torch.randn(
+            (B, 16, 7), device=global_cond.device
+        )
+
+        # 3. 逐步去噪循环 (Reverse Diffusion)
+        self.noise_scheduler.set_timesteps(10) # 推理时可以用更少步加速 (如 DDIM)
+        
+        for t in self.noise_scheduler.timesteps:
+            # 调整输入格式适应 1D CNN
+            model_input = current_actions.permute(0, 2, 1)
             
-#         Returns:
-#             Training: Loss scalar
-#             Inference: Predicted actions
-#         """
-        
-#         B, T, C, H, W = images.shape
-        
-#         # 1. 图像编码 (Flatten time dimension for efficiency)
-#         # [B*T, C, H, W] -> [B*T, Vision_Dim]
-#         flat_images = images.view(B * T, C, H, W)
-#         vision_feats_dict = self.vlm_backbone(flat_images)
-#         raw_img_emb = vision_feats_dict['image_embeds'] # [B*T, Vision_Dim]
-        
-#         # 投影并还原时间维度 -> [B, T, Embed_Dim]
-#         img_emb = self.img_projector(raw_img_emb)
-#         img_emb = img_emb.view(B, T, -1)
-        
-#         # 2. 状态编码
-#         state_emb = self.state_encoder(state) # [B, T, Embed_Dim]
+            # 预测噪声
+            noise_pred = self.noise_pred_net(
+                sample=model_input, 
+                timestep=t, 
+                global_cond=global_cond
+            )
+            # noise_pred = noise_pred.permute(0, 2, 1)
 
-#         # 3. 特征融合 (这里做一个简单的 Early Fusion 示例)
-#         # 将图像特征和状态特征在特征维度拼接，或在时间维度拼接
-#         # 假设我们只用最近的一帧图像作为 Context，或者将所有历史特征作为 Context
-#         # 这里演示：Context = (Image_History + State_History)
-#         # [B, T, Embed] + [B, T, Embed] -> [B, 2*T, Embed] (Concat on time)
-#         context = torch.cat([img_emb, state_emb], dim=1) 
-        
-#         # 4. Action Head 分支
-#         if actions is not None:
-#             # --- Training Mode ---
-#             # 必须返回 Loss
-#             return self.action_head.compute_loss(context, actions)
-#         else:
-#             # --- Inference Mode ---
-#             # 必须返回预测的动作序列
-#             return self.action_head.predict_action(context)
+            # 移除噪声，更新 current_actions
+            current_actions = self.noise_scheduler.step(
+                noise_pred, t, current_actions
+            ).prev_sample
+
+        # 4. 输出最终动作序列
+        return current_actions # 返回去噪后的干净动作

roboimi/vla/conf/agent/resnet_diffusion.yaml

@@ -0,0 +1,22 @@
+# @package agent
+_target_: roboimi.vla.agent.VLAAgent
+
+# Vision Backbone: ResNet-18 with SpatialSoftmax
+vision_backbone:
+  _target_: roboimi.vla.models.backbones.resnet.ResNetBackbone
+  model_name: "microsoft/resnet-18"
+  freeze: true
+
+# Action and Observation Dimensions
+action_dim: 16  # Robot action dimension
+obs_dim: 16     # Proprioception dimension (qpos)
+
+# Prediction Horizons
+pred_horizon: 16   # How many future actions to predict
+obs_horizon: 2     # How many historical observations to use
+
+# Diffusion Parameters
+diffusion_steps: 100  # Number of diffusion timesteps for training
+
+# Camera Configuration
+num_cams: 2  # Number of cameras (e.g., r_vis, top)

roboimi/vla/conf/backbone/resnet.yaml

@@ -0,0 +1,10 @@
+# @package agent.backbone
+_target_: roboimi.vla.models.backbones.resnet.ResNetBackbone
+
+model_name: "microsoft/resnet-18"
+freeze: true
+
+# Output dimension calculation:
+# ResNet-18 final layer has 512 channels
+# After SpatialSoftmax: 512 * 2 = 1024 (x,y coordinates per channel)
+# output_dim: 1024

roboimi/vla/conf/config.yaml

@@ -1,12 +1,13 @@
 defaults:
   - _self_
-  - agent: base_siglip
-  - data: custom_hdf5  # 新增这一行，激活数据配置
+  - agent: resnet_diffusion
+  - data: resnet_dataset
 
 train:
-  batch_size: 4       # 减小 batch size 方便调试
-  lr: 1e-4
-  max_steps: 10
-  log_freq: 10
-  device: "cpu"
-  num_workers: 0      # 调试设为0，验证通过后改为 2 或 4
+  batch_size: 8       # Batch size for training
+  lr: 1e-4            # Learning rate
+  max_steps: 10000    # Maximum training steps
+  log_freq: 100       # Log frequency (steps)
+  save_freq: 1000     # Save checkpoint frequency (steps)
+  device: "cuda"      # Device: "cuda" or "cpu"
+  num_workers: 8      # DataLoader workers (set to 0 for debugging, 8 for production)

roboimi/vla/conf/data/resnet_dataset.yaml

@@ -0,0 +1,18 @@
+# @package data
+_target_: roboimi.vla.data.dataset.RobotDiffusionDataset
+
+# Dataset Directory (CHANGE THIS TO YOUR DATA PATH)
+dataset_dir: "roboimi/demos/dataset/sim_transfer"  # Path to your dataset directory
+
+# Horizon Parameters
+pred_horizon: 16  # Prediction horizon (matches agent.pred_horizon)
+obs_horizon: 2    # Observation horizon (matches agent.obs_horizon)
+action_horizon: 8 # Action execution horizon (used during evaluation)
+
+# Camera Names (CHANGE THIS TO MATCH YOUR CAMERAS)
+camera_names:
+  - r_vis
+  - top
+
+# Normalization Type: 'gaussian' (mean/std) or 'min_max' ([-1, 1])
+normalization_type: gaussian

roboimi/vla/data/dataset.py

@@ -90,52 +90,48 @@ class RobotDiffusionDataset(Dataset):
             # 如果 obs_horizon=2, current_ts=0 -> indices=[0, 0] (Padding)
             # 如果 obs_horizon=2, current_ts=5 -> indices=[4, 5]
             
-            indices = []
-            for i in range(self.obs_horizon):
-                # t - (To - 1) + i
-                query_ts = start_ts - (self.obs_horizon - 1) + i
-                # 边界处理 (Padding first frame)
-                query_ts = max(query_ts, 0)
-                indices.append(query_ts)
-                
-            # 读取 qpos (proprioception)
-            qpos_data = root['observations/qpos']
-            qpos = qpos_data[indices] # smart indexing
-            if self.stats:
-                qpos = self._normalize_data(qpos, self.stats['qpos'])
+            start_idx_raw = start_ts - (self.obs_horizon - 1)
+            start_idx = max(start_idx_raw, 0)
+            end_idx = start_ts + 1
+            pad_len = max(0, -start_idx_raw)
 
-            # 读取 Images
-            # 你有三个视角: angle, r_vis, top
-            # 建议将它们分开返回，或者在 Dataset 里 Concat
+            # Qpos
+            qpos_data = root['observations/qpos']
+            qpos_val = qpos_data[start_idx:end_idx]
+            
+            if pad_len > 0:
+                first_frame = qpos_val[0]
+                padding = np.repeat(first_frame[np.newaxis, :], pad_len, axis=0)
+                qpos_val = np.concatenate([padding, qpos_val], axis=0)
+                
+            if self.stats:
+                qpos_val = self._normalize_data(qpos_val, self.stats['qpos'])
+
+            # Images
             image_dict = {}
             for cam_name in self.camera_names:
-                # HDF5 dataset
                 img_dset = root['observations']['images'][cam_name]
+                imgs_np = img_dset[start_idx:end_idx]  # (T, H, W, C)
                 
-                imgs = []
-                for t in indices:
-                    img = img_dset[t] # (480, 640, 3) uint8
-                    img = torch.from_numpy(img).permute(2, 0, 1).float() / 255.0  # (C, H, W)
-                    imgs.append(img)
+                if pad_len > 0:
+                    first_frame = imgs_np[0]
+                    padding = np.repeat(first_frame[np.newaxis, ...], pad_len, axis=0)
+                    imgs_np = np.concatenate([padding, imgs_np], axis=0)
                 
-                # Stack time dimension: (obs_horizon, 3, H, W)
-                image_dict[cam_name] = torch.stack(imgs)
+                # 转换为 Tensor: (T, H, W, C) -> (T, C, H, W)
+                imgs_tensor = torch.from_numpy(imgs_np).float() / 255.0
+                imgs_tensor = torch.einsum('thwc->tchw', imgs_tensor)
+                image_dict[cam_name] = imgs_tensor
 
-            # -----------------------------
-            # 4. 组装 Batch
-            # -----------------------------
+            # ==============================
+            # 3. 组装 Batch
+            # ==============================
             data_batch = {
-                'action': torch.from_numpy(actions).float(), # (Tp, 16)
-                'qpos': torch.from_numpy(qpos).float(),      # (To, 16)
+                'action': torch.from_numpy(actions).float(),
+                'qpos': torch.from_numpy(qpos_val).float(),
             }
-            # 将图像放入 batch
             for cam_name, img_tensor in image_dict.items():
-                data_batch[f'image_{cam_name}'] = img_tensor # (To, 3, H, W)
-
-            # TODO: 添加 Language Instruction
-            # 如果所有 episode 共享任务，这里可以是固定 embedding
-            # 如果每个 episode 任务不同，你需要一个额外的 meta json 来映射 file_path -> text
-            # data_batch['lang_text'] = "pick up the red cube" 
+                data_batch[f'image_{cam_name}'] = img_tensor
 
             return data_batch
 

roboimi/vla/models/backbones/__init__.py

@@ -1,9 +1,10 @@
 # Backbone models
 from .siglip import SigLIPBackbone
+from .resnet import ResNetBackbone
 # from .clip import CLIPBackbone
 # from .dinov2 import DinoV2Backbone
 
-__all__ = ["SigLIPBackbone"]
+__all__ = ["SigLIPBackbone", "ResNetBackbone"]
 
 # from .debug import DebugBackbone
 # __all__ = ["DebugBackbone"]

roboimi/vla/models/backbones/clip.py

@@ -1 +0,0 @@
-# CLIP Backbone 实现

roboimi/vla/models/backbones/debug.py

@@ -1,30 +0,0 @@
-import torch
-import torch.nn as nn
-from typing import Dict
-from roboimi.vla.core.interfaces import VLABackbone
-
-class DebugBackbone(VLABackbone):
-    """
-    A fake backbone that outputs random tensors.
-    """
-    def __init__(self, embed_dim: int = 768, seq_len: int = 10):
-        super().__init__()
-        self._embed_dim = embed_dim
-        self.seq_len = seq_len
-        # A dummy trainable parameter
-        self.dummy_param = nn.Parameter(torch.zeros(1))
-
-    def forward(self, obs: Dict[str, torch.Tensor]) -> torch.Tensor:
-        batch_size = obs['image'].shape[0]
-        
-        # 1. Generate random noise
-        noise = torch.randn(batch_size, self.seq_len, self._embed_dim, device=obs['image'].device)
-        
-        # 2. CRITICAL FIX: Add the dummy parameter to the noise.
-        # This connects 'noise' to 'self.dummy_param' in the computation graph.
-        # The value doesn't change (since param is 0), but the gradient path is established.
-        return noise + self.dummy_param
-
-    @property
-    def embed_dim(self) -> int:
-        return self._embed_dim

roboimi/vla/models/backbones/dinov2.py

@@ -1 +0,0 @@
-# DinoV2 Backbone 实现

roboimi/vla/models/backbones/resnet.py

@@ -0,0 +1,83 @@
+from roboimi.vla.core.interfaces import VLABackbone
+from transformers import ResNetModel
+from torchvision import transforms
+import torch
+import torch.nn as nn
+
+class ResNetBackbone(VLABackbone):
+    def __init__(
+            self,
+            model_name = "microsoft/resnet-18",
+            freeze: bool = True,
+    ):
+        super().__init__()
+        self.model = ResNetModel.from_pretrained(model_name)
+        self.out_channels = self.model.config.hidden_sizes[-1]
+        self.transform = transforms.Compose([
+            transforms.Resize((384, 384)),
+            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+        ])
+        self.spatial_softmax = SpatialSoftmax(num_rows=12, num_cols=12)
+        if freeze:
+            self._freeze_parameters()
+
+    def _freeze_parameters(self):
+        print("❄️ Freezing ResNet Backbone parameters")
+        for param in self.model.parameters():
+            param.requires_grad = False
+        self.model.eval()
+
+    def forward_single_image(self, image):
+        B, T, C, H, W = image.shape
+        image = image.view(B * T, C, H, W)
+        image = self.transform(image)
+        feature_map = self.model(image).last_hidden_state  # (B*T, D, H', W')
+        features = self.spatial_softmax(feature_map)  # (B*T, D*2)
+        return features
+    
+    def forward(self, images):
+        any_tensor = next(iter(images.values()))
+        B, T = any_tensor.shape[:2]
+        features_all = []
+        sorted_cam_names = sorted(images.keys())
+        for cam_name in sorted_cam_names:
+            img = images[cam_name]
+            features = self.forward_single_image(img)  # (B*T, D*2)
+            features_all.append(features)
+        combined_features = torch.cat(features_all, dim=1)  # (B*T, Num_Cams*D*2)
+        return combined_features.view(B, T, -1)
+
+    @property
+    def output_dim(self):
+        """Output dimension after spatial softmax: out_channels * 2"""
+        return self.out_channels * 2
+
+class SpatialSoftmax(nn.Module):
+    """
+    将特征图 (N, C, H, W) 转换为坐标特征 (N, C*2)
+    """
+    def __init__(self, num_rows, num_cols, temperature=None):
+        super().__init__()
+        self.temperature = nn.Parameter(torch.ones(1))
+        # 创建网格坐标
+        pos_x, pos_y = torch.meshgrid(
+            torch.linspace(-1, 1, num_rows),
+            torch.linspace(-1, 1, num_cols),
+            indexing='ij'
+        )
+        self.register_buffer('pos_x', pos_x.reshape(-1))
+        self.register_buffer('pos_y', pos_y.reshape(-1))
+
+    def forward(self, x):
+        N, C, H, W = x.shape
+        x = x.view(N, C, -1) # (N, C, H*W)
+        
+        # 计算 Softmax 注意力图
+        softmax_attention = torch.nn.functional.softmax(x / self.temperature, dim=2)
+        
+        # 计算期望坐标 (x, y)
+        expected_x = torch.sum(softmax_attention * self.pos_x, dim=2, keepdim=True)
+        expected_y = torch.sum(softmax_attention * self.pos_y, dim=2, keepdim=True)
+        
+        # 拼接并展平 -> (N, C*2)
+        return torch.cat([expected_x, expected_y], dim=2).reshape(N, -1)

roboimi/vla/models/heads/__init__.py

@@ -1,9 +1,8 @@
 # # Action Head models
-from .diffusion import DiffusionHead
+from .diffusion import ConditionalUnet1D
 # from .act import ACTHead
 
-__all__ = ["DiffusionHead"]
+__all__ = ["ConditionalUnet1D"]
 
 # from .debug import DebugHead
-
 # __all__ = ["DebugHead"]

roboimi/vla/models/heads/act.py

@@ -1 +0,0 @@
-# ACT-VAE Action Head 实现

roboimi/vla/models/heads/debug.py

@@ -1,33 +0,0 @@
-import torch
-import torch.nn as nn
-from typing import Dict, Optional
-from roboimi.vla.core.interfaces import VLAHead
-
-class DebugHead(VLAHead):
-    """
-    A fake Action Head using MSE Loss.
-    Replaces complex Diffusion/ACT policies for architecture verification.
-    """
-    def __init__(self, input_dim: int, action_dim: int, chunk_size: int = 16):
-        super().__init__()
-        # Simple regression from embedding -> action chunk
-        self.regressor = nn.Linear(input_dim, chunk_size * action_dim)
-        self.action_dim = action_dim
-        self.chunk_size = chunk_size
-        self.loss_fn = nn.MSELoss()
-
-    def forward(self, embeddings: torch.Tensor, actions: Optional[torch.Tensor] = None) -> Dict[str, torch.Tensor]:
-        # Simple pooling over sequence dimension to get (B, Hidden)
-        pooled_embed = embeddings.mean(dim=1) 
-        
-        # Predict actions: (B, Chunk * Act_Dim) -> (B, Chunk, Act_Dim)
-        pred_flat = self.regressor(pooled_embed)
-        pred_actions = pred_flat.view(-1, self.chunk_size, self.action_dim)
-
-        output = {"pred_actions": pred_actions}
-
-        if actions is not None:
-            # Calculate MSE Loss against ground truth
-            output["loss"] = self.loss_fn(pred_actions, actions)
-            
-        return output

roboimi/vla/models/heads/diffusion.py

@@ -5,170 +5,290 @@ from typing import Dict, Optional
 from diffusers import DDPMScheduler
 from roboimi.vla.core.interfaces import VLAHead
 
-class DiffusionHead(VLAHead):
-    def __init__(
-        self,
-        input_dim: int,      # 来自 Projector 的维度 (e.g. 384)
-        action_dim: int,     # 动作维度 (e.g. 16)
-        chunk_size: int,     # 预测视界 (e.g. 16)
-        n_timesteps: int = 100, # 扩散步数
-        hidden_dim: int = 256
-    ):
+from typing import Union
+import logging
+import torch
+import torch.nn as nn
+import einops
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops.layers.torch import Rearrange
+import math
+
+
+class SinusoidalPosEmb(nn.Module):
+    def __init__(self, dim):
         super().__init__()
-        self.action_dim = action_dim
-        self.chunk_size = chunk_size
-        
-        # 1. 噪声调度器 (DDPM)
-        self.scheduler = DDPMScheduler(
-            num_train_timesteps=n_timesteps,
-            beta_schedule='squaredcos_cap_v2', # 现代 Diffusion 常用调度
-            clip_sample=True,
-            prediction_type='epsilon' # 预测噪声
-        )
+        self.dim = dim
 
-        # 2. 噪声预测网络 (Noise Predictor Network)
-        # 输入: Noisy Action + Time Embedding + Image Embedding
-        # 这是一个简单的 Conditional MLP/ResNet 结构
-        self.time_emb = nn.Sequential(
-            nn.Linear(1, hidden_dim),
+    def forward(self, x):
+        device = x.device
+        half_dim = self.dim // 2
+        emb = math.log(10000) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, device=device) * -emb)
+        emb = x[:, None] * emb[None, :]
+        emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
+        return emb
+
+class Downsample1d(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.conv = nn.Conv1d(dim, dim, 3, 2, 1)
+
+    def forward(self, x):
+        return self.conv(x)
+
+class Upsample1d(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.conv = nn.ConvTranspose1d(dim, dim, 4, 2, 1)
+
+    def forward(self, x):
+        return self.conv(x)
+
+class Conv1dBlock(nn.Module):
+    '''
+        Conv1d --> GroupNorm --> Mish
+    '''
+
+    def __init__(self, inp_channels, out_channels, kernel_size, n_groups=8):
+        super().__init__()
+
+        self.block = nn.Sequential(
+            nn.Conv1d(inp_channels, out_channels, kernel_size, padding=kernel_size // 2),
+            # Rearrange('batch channels horizon -> batch channels 1 horizon'),
+            nn.GroupNorm(n_groups, out_channels),
+            # Rearrange('batch channels 1 horizon -> batch channels horizon'),
             nn.Mish(),
-            nn.Linear(hidden_dim, hidden_dim)
         )
-        
-        self.cond_proj = nn.Linear(input_dim, hidden_dim) # 把图像特征投影一下
-        
-        # 主干网络 (由几个 Residual Block 组成)
-        self.mid_layers = nn.ModuleList([
-            nn.Sequential(
-                nn.Linear(hidden_dim + action_dim * chunk_size, hidden_dim),
-                nn.LayerNorm(hidden_dim),
-                nn.Mish(),
-                nn.Linear(hidden_dim, hidden_dim + action_dim * chunk_size) # 简单的残差
-            ) for _ in range(3)
+
+    def forward(self, x):
+        return self.block(x)
+
+class ConditionalResidualBlock1D(nn.Module):
+    def __init__(self, 
+            in_channels, 
+            out_channels, 
+            cond_dim,
+            kernel_size=3,
+            n_groups=8,
+            cond_predict_scale=False):
+        super().__init__()
+        self.blocks = nn.ModuleList([
+            Conv1dBlock(in_channels, out_channels, kernel_size, n_groups=n_groups),
+            Conv1dBlock(out_channels, out_channels, kernel_size, n_groups=n_groups),
         ])
-        
-        # 输出层: 预测噪声 (Shape 与 Action 相同)
-        self.final_layer = nn.Linear(hidden_dim + action_dim * chunk_size, action_dim * chunk_size)
 
-    def forward(self, embeddings: torch.Tensor, actions: Optional[torch.Tensor] = None) -> Dict[str, torch.Tensor]:
-        """
-        Unified interface for Training and Inference.
-        """
-        device = embeddings.device
-        
-        # --- 1. 处理条件 (Conditioning) ---
-        # embeddings: (B, Seq, Dim). 我们这里做一个简化，做 Average Pooling 变成 (B, Dim)
-        # 如果你想做更复杂的 Cross-Attention，可以在这里改
-        global_cond = embeddings.mean(dim=1) 
-        cond_feat = self.cond_proj(global_cond) # (B, Hidden)
 
-        # =========================================
-        # 分支 A: 训练模式 (Training)
-        # =========================================
-        if actions is not None:
-            batch_size = actions.shape[0]
-            
-            # 1.1 准备数据 (Flatten: B, Chunk, ActDim -> B, Chunk*ActDim)
-            actions_flat = actions.view(batch_size, -1)
-            
-            # 1.2 采样噪声和时间步
-            noise = torch.randn_like(actions_flat)
-            timesteps = torch.randint(
-                0, self.scheduler.config.num_train_timesteps, 
-                (batch_size,), device=device
-            ).long()
-            
-            # 1.3 加噪 (Forward Diffusion)
-            noisy_actions = self.scheduler.add_noise(actions_flat, noise, timesteps)
-            
-            # 1.4 预测噪声 (Network Forward)
-            pred_noise = self._predict_noise(noisy_actions, timesteps, cond_feat)
-            
-            # 1.5 计算 Loss (MSE between actual noise and predicted noise)
-            loss = nn.functional.mse_loss(pred_noise, noise)
-            
-            return {"loss": loss}
 
-        # =========================================
-        # 分支 B: 推理模式 (Inference)
-        # =========================================
+        cond_channels = out_channels
+        if cond_predict_scale:
+            cond_channels = out_channels * 2
+        self.cond_predict_scale = cond_predict_scale
+        self.out_channels = out_channels
+        self.cond_encoder = nn.Sequential(
+            nn.Mish(),
+            nn.Linear(cond_dim, cond_channels),
+            Rearrange('batch t -> batch t 1'),
+        )
+
+        # make sure dimensions compatible
+        self.residual_conv = nn.Conv1d(in_channels, out_channels, 1) \
+            if in_channels != out_channels else nn.Identity()
+
+    def forward(self, x, cond):
+        '''
+            x : [ batch_size x in_channels x horizon ]
+            cond : [ batch_size x cond_dim]
+
+            returns:
+            out : [ batch_size x out_channels x horizon ]
+        '''
+        out = self.blocks[0](x)
+        embed = self.cond_encoder(cond)
+        if self.cond_predict_scale:
+            embed = embed.reshape(
+                embed.shape[0], 2, self.out_channels, 1)
+            scale = embed[:,0,...]
+            bias = embed[:,1,...]
+            out = scale * out + bias
         else:
-            batch_size = embeddings.shape[0]
-            
-            # 2.1 从纯高斯噪声开始
-            noisy_actions = torch.randn(
-                batch_size, self.chunk_size * self.action_dim, 
-                device=device
-            )
-            
-            # 2.2 逐步去噪 (Reverse Diffusion Loop)
-            # 使用 scheduler.timesteps 自动处理步长
-            self.scheduler.set_timesteps(self.scheduler.config.num_train_timesteps)
-            
-            for t in self.scheduler.timesteps:
-                # 构造 batch 的 t
-                timesteps = torch.tensor([t], device=device).repeat(batch_size)
-                
-                # 预测噪声
-                # 注意：diffusers 的 step 需要 model_output
-                model_output = self._predict_noise(noisy_actions, timesteps, cond_feat)
-                
-                # 移除噪声 (Step)
-                noisy_actions = self.scheduler.step(
-                    model_output, t, noisy_actions
-                ).prev_sample
+            out = out + embed
+        out = self.blocks[1](out)
+        out = out + self.residual_conv(x)
+        return out
 
-            # 2.3 Reshape 回 (B, Chunk, ActDim)
-            pred_actions = noisy_actions.view(batch_size, self.chunk_size, self.action_dim)
-            
-            return {"pred_actions": pred_actions}
 
-    def _predict_noise(self, noisy_actions, timesteps, cond_feat):
-        """内部辅助函数：运行简单的 MLP 网络"""
-        # Time Embed
-        t_emb = self.time_emb(timesteps.float().unsqueeze(-1)) # (B, Hidden)
+class ConditionalUnet1D(nn.Module):
+    def __init__(self, 
+        input_dim,
+        local_cond_dim=None,
+        global_cond_dim=None,
+        diffusion_step_embed_dim=256,
+        down_dims=[256,512,1024],
+        kernel_size=3,
+        n_groups=8,
+        cond_predict_scale=False
+        ):
+        super().__init__()
+        all_dims = [input_dim] + list(down_dims)
+        start_dim = down_dims[0]
+
+        dsed = diffusion_step_embed_dim
+        diffusion_step_encoder = nn.Sequential(
+            SinusoidalPosEmb(dsed),
+            nn.Linear(dsed, dsed * 4),
+            nn.Mish(),
+            nn.Linear(dsed * 4, dsed),
+        )
+        cond_dim = dsed
+        if global_cond_dim is not None:
+            cond_dim += global_cond_dim
+
+        in_out = list(zip(all_dims[:-1], all_dims[1:]))
+
+        local_cond_encoder = None
+        if local_cond_dim is not None:
+            _, dim_out = in_out[0]
+            dim_in = local_cond_dim
+            local_cond_encoder = nn.ModuleList([
+                # down encoder
+                ConditionalResidualBlock1D(
+                    dim_in, dim_out, cond_dim=cond_dim, 
+                    kernel_size=kernel_size, n_groups=n_groups,
+                    cond_predict_scale=cond_predict_scale),
+                # up encoder
+                ConditionalResidualBlock1D(
+                    dim_in, dim_out, cond_dim=cond_dim, 
+                    kernel_size=kernel_size, n_groups=n_groups,
+                    cond_predict_scale=cond_predict_scale)
+            ])
+
+        mid_dim = all_dims[-1]
+        self.mid_modules = nn.ModuleList([
+            ConditionalResidualBlock1D(
+                mid_dim, mid_dim, cond_dim=cond_dim,
+                kernel_size=kernel_size, n_groups=n_groups,
+                cond_predict_scale=cond_predict_scale
+            ),
+            ConditionalResidualBlock1D(
+                mid_dim, mid_dim, cond_dim=cond_dim,
+                kernel_size=kernel_size, n_groups=n_groups,
+                cond_predict_scale=cond_predict_scale
+            ),
+        ])
+
+        down_modules = nn.ModuleList([])
+        for ind, (dim_in, dim_out) in enumerate(in_out):
+            is_last = ind >= (len(in_out) - 1)
+            down_modules.append(nn.ModuleList([
+                ConditionalResidualBlock1D(
+                    dim_in, dim_out, cond_dim=cond_dim, 
+                    kernel_size=kernel_size, n_groups=n_groups,
+                    cond_predict_scale=cond_predict_scale),
+                ConditionalResidualBlock1D(
+                    dim_out, dim_out, cond_dim=cond_dim, 
+                    kernel_size=kernel_size, n_groups=n_groups,
+                    cond_predict_scale=cond_predict_scale),
+                Downsample1d(dim_out) if not is_last else nn.Identity()
+            ]))
+
+        up_modules = nn.ModuleList([])
+        for ind, (dim_in, dim_out) in enumerate(reversed(in_out[1:])):
+            is_last = ind >= (len(in_out) - 1)
+            up_modules.append(nn.ModuleList([
+                ConditionalResidualBlock1D(
+                    dim_out*2, dim_in, cond_dim=cond_dim,
+                    kernel_size=kernel_size, n_groups=n_groups,
+                    cond_predict_scale=cond_predict_scale),
+                ConditionalResidualBlock1D(
+                    dim_in, dim_in, cond_dim=cond_dim,
+                    kernel_size=kernel_size, n_groups=n_groups,
+                    cond_predict_scale=cond_predict_scale),
+                Upsample1d(dim_in) if not is_last else nn.Identity()
+            ]))
         
-        # Fusion: Concat Action + (Condition * Time) 
-        # 这里用简单的相加融合，实际可以更复杂
-        fused_feat = cond_feat + t_emb
+        final_conv = nn.Sequential(
+            Conv1dBlock(start_dim, start_dim, kernel_size=kernel_size),
+            nn.Conv1d(start_dim, input_dim, 1),
+        )
+
+        self.diffusion_step_encoder = diffusion_step_encoder
+        self.local_cond_encoder = local_cond_encoder
+        self.up_modules = up_modules
+        self.down_modules = down_modules
+        self.final_conv = final_conv
+
+
+    def forward(self, 
+            sample: torch.Tensor, 
+            timestep: Union[torch.Tensor, float, int], 
+            local_cond=None, global_cond=None, **kwargs):
+        """
+        x: (B,T,input_dim)
+        timestep: (B,) or int, diffusion step
+        local_cond: (B,T,local_cond_dim)
+        global_cond: (B,global_cond_dim)
+        output: (B,T,input_dim)
+        """
+        sample = einops.rearrange(sample, 'b h t -> b t h')
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            timesteps = torch.tensor([timesteps], dtype=torch.long, device=sample.device)
+        elif torch.is_tensor(timesteps) and len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        global_feature = self.diffusion_step_encoder(timesteps)
+
+        if global_cond is not None:
+            global_feature = torch.cat([
+                global_feature, global_cond
+            ], axis=-1)
         
-        # Concat input
-        x = torch.cat([noisy_actions, fused_feat], dim=-1) # 注意这里维度需要对齐，或者用 MLP 映射
+        # encode local features
+        h_local = list()
+        if local_cond is not None:
+            local_cond = einops.rearrange(local_cond, 'b h t -> b t h')
+            resnet, resnet2 = self.local_cond_encoder
+            x = resnet(local_cond, global_feature)
+            h_local.append(x)
+            x = resnet2(local_cond, global_feature)
+            h_local.append(x)
         
-        # 修正：上面的 concat 维度可能不对，为了简化代码，我们用一种更简单的方式：
-        # 将 cond_feat 加到 input 里需要维度匹配。
-        # 这里重写一个极简的 Forward:
-        
-        # 正确做法：先将 x 映射到 hidden，再加 t_emb 和 cond_feat
-        # 但为了复用 self.mid_layers 定义的 Linear(Hidden + Input)...
-        # 我们用最傻瓜的方式：Input = Action，Condition 直接拼接到每一层或者只拼输入
-        
-        # 让我们修正一下网络结构逻辑，确保不报错：
-        # Input: NoisyAction (Dim_A)
-        # Cond:  Hidden (Dim_H)
-        
-        # 这种临时写的 MLP 容易维度不匹配，我们改用一个极其稳健的计算流：
-        # x = Action
-        # h = Cond + Time
-        # input = cat([x, h]) -> Linear -> Output
-        
-        # 重新定义 _predict_noise 的逻辑依赖于 __init__ 里的定义。
-        # 为了保证一次跑通，我使用动态 cat:
-        
-        x = noisy_actions
-        # 假设 mid_layers 的输入是 hidden_dim + action_flat_dim
-        # 我们把 condition 映射成 hidden_dim，然后 concat
-        
-        # 真正的计算流：
-        h = cond_feat + t_emb # (B, Hidden)
-        
-        # 把 h 拼接到 x 上 (前提是 x 是 action flat)
-        # Linear 输入维度是 Hidden + ActFlat
-        model_input = torch.cat([h, x], dim=-1) 
-        
-        for layer in self.mid_layers:
-            # Residual connection mechanism
-            out = layer(model_input)
-            model_input = out + model_input # Simple ResNet
-            
-        return self.final_layer(model_input)
+        x = sample
+        h = []
+        for idx, (resnet, resnet2, downsample) in enumerate(self.down_modules):
+            x = resnet(x, global_feature)
+            if idx == 0 and len(h_local) > 0:
+                x = x + h_local[0]
+            x = resnet2(x, global_feature)
+            h.append(x)
+            x = downsample(x)
+
+        for mid_module in self.mid_modules:
+            x = mid_module(x, global_feature)
+
+        for idx, (resnet, resnet2, upsample) in enumerate(self.up_modules):
+            x = torch.cat((x, h.pop()), dim=1)
+            x = resnet(x, global_feature)
+            # The correct condition should be:
+            # if idx == (len(self.up_modules)-1) and len(h_local) > 0:
+            # However this change will break compatibility with published checkpoints.
+            # Therefore it is left as a comment.
+            if idx == len(self.up_modules) and len(h_local) > 0:
+                x = x + h_local[1]
+            x = resnet2(x, global_feature)
+            x = upsample(x)
+
+        x = self.final_conv(x)
+
+        x = einops.rearrange(x, 'b t h -> b h t')
+        return x
+