MetaWorld-VLA-openai-clip-vit

MetaWorld-CLIP-BC-VLA

A lightweight Vision-Language-Action (VLA) baseline for MetaWorld robot-arm tasks using a pretrained CLIP-ViT vision transformer(openai/clip-vit-base-patch32), a small text transformer, and robot-state fusion—trained with behavior cloning and evaluated in a MetaWorld MT1 wrapper.

• Vision: pretrained CLIP ViT (transformer) as an image encoder
• Language: small transformer encoder over token IDs
• Robot state: fused with vision+language embeddings
• Policy head: predicts continuous actions (MetaWorld-style)

---

Demo Video

<a href="https://youtu.be/7aWUy5V1rC0"> <img src="assets/frame.jpg" alt="Demo Video" width="500"> </a>

Features

• ✅ Uses a pretrained transformer vision encoder (openai/clip-vit-base-patch32)
• ✅ Works with MetaWorld MT1 tasks
• ✅ Vision-Language-State → Action fusion
• ✅ Minimal training code (BC with MSE + AdamW)
• ✅ Saves checkpoints compatible with the provided test.py rollout script

---

Install dependencies

pip install -r requirements.txt

Project Layout

1. Collect Demonstrations (soccer-v3) This uses MetaWorld’s built-in expert policies (metaworld.policies) to generate demonstrations:

python3 collect_data.py \
--env-name soccer-v3 \
--camera-name corner \
--episodes 100 \
--max-steps 140 \
--output-path data/metaworld_soccer_bc.npz  \
--resize-to 480

1. Train the VLA policy

python3 train.py \
--data data/metaworld_soccer_bc.npz \
--out checkpoints/metaworld_soccer.pt \
--device cuda

1. Testing the model

python3 test.py \
--checkpoint checkpoints/metaworld_soccer.pt \
--env-name soccer-v3 \
--episodes 5 \
--max-steps 150 \
--instruction "Shoot the ball into the goal" \
--device cuda \
--save-video \
--video-dir videos

Model Architecture (Layer-by-Layer)

A lightweight Vision-Language-Action (VLA) policy for MetaWorld trained with Behavior Cloning (BC).

The model class is: models/vla_clip_bc.py::VLAClipBC

At a high level:

(RGB image, instruction token IDs, robot state) → fused embedding → action

---

Overview

Inputs

• img: RGB image tensor (B, 3, H, W) in [0, 1]
• text_ids: token IDs (B, T) from SimpleTokenizer (padding id 0)
• state: robot state vector (B, S)

Outputs

• action: continuous action (B, A) (optionally tanh squashed to [-1, 1])

---

Module Breakdown

1) Vision Branch (Pretrained CLIP ViT)

Implemented in VLAClipBC.encode_vision():

1. Preprocess

   • Resize image to 224×224
F.interpolate(img, size=(224,224))
   • Normalize with CLIP constants
     img = (img - clip_mean) / clip_std

2. CLIP Vision Transformer

   • self.vision = CLIPVisionModel.from_pretrained("openai/clip-vit-base-patch32")
   • Uses out.pooler_output (shape (B, 768) for ViT-B/32)

3. Projection to d_model

   • self.vision_proj = Linear(768 → D)
   • Result: vision embedding v with shape (B, D)
     (default D = 512)

CLIPVisionModel (ViT-B/32) internals (high-level)

• Patch embedding + class token + positional embeddings
• Transformer encoder stack (typical ViT-B/32)
  • ~12 transformer encoder layers
  • Each layer includes:
    • LayerNorm
    • Multi-Head Self-Attention
    • residual connection
    • LayerNorm
    • MLP (Linear → GELU → Linear)
    • residual connection

Note: when freeze_vision=True, CLIP vision parameters are frozen.

---

2) Text Branch (SimpleTextTransformer)

Implemented in SimpleTextTransformer.forward():

1. Embeddings

   • Token embedding: Embedding(vocab_size, D, padding_idx=0)
   • Position embedding: Embedding(max_len=64, D)
   • Sum: x = tok_emb(text_ids) + pos_emb(pos)

2. Transformer Encoder

   • nn.TransformerEncoder(num_layers=text_layers)
   • Each layer (because norm_first=True):
     • LayerNorm → SelfAttention → residual
     • LayerNorm → FFN (Linear → GELU → Linear) → residual
   • Then: LayerNorm(D)

3. Masked Mean Pooling

   • Mask pads where text_ids == 0
   • Average over non-pad tokens
   • Result: text embedding t with shape (B, D)

Default config:

• text_layers = 2
• text_heads = 8
• max_len = 64

---

3) State Branch (Robot State Encoder)

Implemented in VLAClipBC.encode_state():

1. Normalize state

   • state_mean and state_std are stored in buffers (from training dataset stats)
   • s = (state - state_mean) / state_std

2. MLP Encoder

   • LayerNorm(S)
   • Linear(S → 2D)
   • GELU
   • Linear(2D → D)

Result: state embedding s with shape (B, D)

---

4) Fusion Branch (3-token Transformer)

Implemented in VLAClipBC.forward():

1. Stack 3 tokens

   • tokens = stack([v, t, s], dim=1)
   • Shape: (B, 3, D)
     token0 = vision, token1 = text, token2 = state

2. Fusion TransformerEncoder

   • nn.TransformerEncoder(num_layers=fusion_layers)
   • Each layer:
     • LayerNorm → SelfAttention → residual
     • LayerNorm → FFN (Linear → GELU → Linear) → residual

3. Take token 0

   • x = LayerNorm(fused[:, 0])
   • Shape: (B, D)

Default config:

• fusion_layers = 2
• fusion_heads = 8

---

5) Action Head (Policy Head)

Implemented in VLAClipBC.forward():

• Linear(D → 2D)
• GELU
• Linear(2D → A)
• Optional: tanh(action) (keeps actions in [-1, 1])

Result: (B, A)

---

Shape Flow (End-to-End)

Assuming default D=512 and CLIP ViT-B/32 hidden size 768:

• Image: (B,3,H,W) → resize/normalize → CLIP ViT → (B,768) → vision_proj → v (B,512)
• Text: (B,T) → embeddings → text transformer → masked mean pool → t (B,512)
• State: (B,S) → normalize → MLP → s (B,512)
• Stack tokens: (B,3,512) → fusion transformer → take token0 → (B,512)
• Head MLP: (B,512) → (B,A) → optional tanh

---

Training Objective (Behavior Cloning)

Training uses Behavior Cloning (BC) from demonstrations:

• Dataset provides (img, text_ids, state) → expert_action
• Loss:
  • MSE(pred_action, expert_action)
• Optimizer:
  • AdamW(trainable_params, lr, weight_decay)
• Optional:
  • Gradient clipping (e.g., clip_grad_norm_)

---

How to Print / Inspect All Layers

from models.vla_clip_bc import VLAClipBC

model = VLAClipBC(vocab_size=2000, state_dim=39, action_dim=4, d_model=512, freeze_vision=True)
print(model)           # prints full nested module tree
print(model.vision)    # prints CLIP vision internals

Contributing

Contributions are welcome. If you find a bug or have a feature request, please open an issue or submit a pull request.

License

This project is released under the <b>MIT License</b>. Feel free to use, modify, and distribute for your own projects.