跳到主要内容

pose_transformer

Module: GBC.utils.data_preparation.pose_transformer

This module provides neural network architectures for mapping SMPL pose representations to humanoid robot joint actions. The core PoseTransformer class implements a Vision Transformer (ViT) inspired approach that achieves superior performance in pose-to-action translation tasks.

📚 Dependencies​

This module requires the following Python packages:

  • torch - PyTorch framework for neural networks
  • torch.nn - Neural network layers and functions
  • torch.nn.functional - Functional interface for operations
  • typing - Type annotations and hints

🏗️ Core Architecture​

🤖 PoseTransformer​

Module Name: GBC.utils.data_preparation.pose_transformer.PoseTransformer

Definition:

class PoseTransformer(nn.Module):
    def __init__(self, 
                 load_hands: bool = False,
                 num_actions: int = 29,
                 embedding_dim: int = 64,
                 joint_embedding_dim: int = 256,
                 num_heads: int = 4,
                 num_layers: int = 4
                ):

📥 Input Parameters:

  • load_hands (bool): Whether to include hand joints (21 vs 51 joints). Default: False
  • num_actions (int): Number of output robot DOF actions. Default: 29
  • embedding_dim (int): Dimension of feature embeddings. Default: 64
  • joint_embedding_dim (int): Dimension of joint space embedding. Default: 256
  • num_heads (int): Number of attention heads in transformer. Default: 4
  • num_layers (int): Number of transformer encoder layers. Default: 4

🔧 Functionality: Neural network for converting SMPL pose parameters (angle-axis representation) to humanoid robot joint actions through inverse kinematics learning.

🧠 Design Philosophy - ViT-Inspired Architecture​

🎯 Core Innovation: Joint Tokenization Strategy​

The PoseTransformer employs a revolutionary approach inspired by Vision Transformers (ViT) that treats pose conversion as a sequence modeling problem with joint tokenization:

🔬 Key Design Principle: Instead of treating pose data as a monolithic vector, the model partitions joint motions into discrete tokens and projects them into a high-dimensional latent space to extract intrinsic joint relationships and movement patterns.

📊 Architecture Components​

1. Joint Space Tokenization​

x = x.view(N, self.num_joints, self.input_dim)  # Reshape to joint tokens
x = x.permute(0, 2, 1)                         # Prepare for embedding
x = self.joint_embedding(x)                    # Project to joint space

🔧 Functionality:

  • Input Reshaping: (N, num_joints * 3) → (N, num_joints, 3)
  • Joint Tokenization: Each joint's 3D angle-axis becomes a token
  • Spatial Projection: Joint dimensions mapped to high-dimensional space

2. Feature Embedding & Positional Encoding​

x = self.embedding(x)              # Map to embedding dimension
x += self.positional_encoding      # Add learnable position information

🔧 Functionality:

  • Feature Mapping: Project joint tokens to transformer-compatible dimensions
  • Positional Awareness: Learnable encoding preserves joint ordering and relationships

3. Transformer Processing​

x = self.transformer(x)            # Multi-head self-attention processing
x = x.mean(dim=1)                 # Global feature aggregation

🔧 Functionality:

  • Self-Attention: Captures complex inter-joint dependencies
  • Multi-Head Processing: Parallel attention patterns for different motion aspects
  • Global Aggregation: Mean pooling for final pose representation

4. Action Decoding​

x = self.decoder(x)               # Map to robot action space

🔧 Functionality:

  • Dimensionality Mapping: Transform global features to robot DOF actions
  • Action Space Conversion: SMPL angle-axis → Robot joint commands

⚡ Superior Performance Analysis​

🏆 Benchmark Comparisons​

The ViT-inspired joint tokenization approach demonstrates exceptional performance advantages over traditional architectures:

📈 Performance Hierarchy (Best to Worst):

  1. 🥇 PoseTransformer (ViT-inspired) - Joint tokenization with transformer processing
  2. 🥈 Standard Transformers - CLS token-based prediction transformers
  3. 🥉 Multi-Layer Perceptrons (MLPs) - Deep feedforward networks
  4. 🏃 Convolutional Neural Networks (CNNs) - Spatial feature extraction approaches
  5. 📉 Other Network Structures - Alternative pose processing architectures

🔬 Why ViT Design Excels​

✨ Joint Relationship Modeling​

Traditional Approach Problems:

  • Treats pose as flat vector, losing joint structure
  • Limited capacity for modeling complex joint interactions
  • Poor generalization to novel pose configurations

ViT-Inspired Solution:

  • Preserves joint structure through tokenization
  • Captures complex dependencies via self-attention
  • Learns intrinsic movement patterns across joint sequences

🧩 Attention Mechanism Advantages​

# Multi-head attention captures diverse joint relationships
encoder = nn.TransformerEncoderLayer(d_model=embedding_dim, nhead=num_heads)
self.transformer = nn.TransformerEncoder(encoder, num_layers=num_layers)

🔧 Benefits:

  • Parallel Processing: Multiple attention heads capture different interaction patterns
  • Long-Range Dependencies: Direct connections between any joint pairs
  • Adaptive Focus: Dynamic attention weights based on pose context

🎯 Latent Space Projection Benefits​

self.joint_embedding = nn.Linear(self.num_joints, joint_embedding_dim)
self.embedding = nn.Linear(self.input_dim, embedding_dim)

🔧 Advantages:

  • Rich Representation: High-dimensional space captures subtle joint features
  • Learned Features: Automatic discovery of optimal joint representations
  • Generalization: Better adaptation to unseen pose variations

🔧 Model Architecture Details​

📊 Input/Output Specifications​

Input Shape: (N, num_joints * 3)

  • N: Batch size
  • num_joints: 21 (body only) or 51 (with hands)
  • 3: Angle-axis representation (x, y, z rotations)

Output Shape: (N, num_actions)

  • N: Batch size
  • num_actions: Robot DOF count (typically 29 for humanoids)

🔄 Forward Pass Pipeline​

def forward(self, x: torch.Tensor) -> torch.Tensor:
    N = x.shape[0]
    
    # 1. Joint Tokenization
    x = x.view(N, self.num_joints, self.input_dim)
    x = x.permute(0, 2, 1)
    x = self.joint_embedding(x)
    
    # 2. Feature Embedding + Positional Encoding
    x = x.permute(0, 2, 1)
    x = self.embedding(x) 
    x += self.positional_encoding
    
    # 3. Transformer Processing
    x = self.transformer(x)
    
    # 4. Global Aggregation + Decoding
    x = x.mean(dim=1)
    x = self.decoder(x)
    return x

⚙️ Configuration Parameters​

Architecture Scaling:

  • Small Model: embedding_dim=64, num_heads=4, num_layers=4
  • Medium Model: embedding_dim=128, num_heads=8, num_layers=6
  • Large Model: embedding_dim=256, num_heads=16, num_layers=8

Joint Configuration:

  • Body Only: load_hands=False (21 joints)
  • Full Body: load_hands=True (51 joints including fingers)

💡 Usage Examples​

🚀 Basic Model Creation​

from GBC.utils.data_preparation.pose_transformer import PoseTransformer

# Create transformer for body-only poses
model = PoseTransformer(
    load_hands=False,      # Body joints only (21 joints)
    num_actions=29,        # Unitree H1 DOF count
    embedding_dim=64,      # Feature dimension
    joint_embedding_dim=256, # Joint space dimension
    num_heads=4,           # Attention heads
    num_layers=4           # Transformer layers
)

🔧 Advanced Configuration​

# High-capacity model for complex robots
advanced_model = PoseTransformer(
    load_hands=True,       # Include hand joints (51 total)
    num_actions=42,        # More complex robot (arms + legs + torso)
    embedding_dim=128,     # Higher feature dimension
    joint_embedding_dim=512, # Richer joint representation
    num_heads=8,           # More attention patterns
    num_layers=6           # Deeper processing
)

Model Persistence​

# Save trained model
torch.save({
    'model_state_dict': model.state_dict(),
    'config': {
        'load_hands': False,
        'num_actions': 29,
        'embedding_dim': 64,
        # ... other config
    }
}, 'pose_transformer.pt')

# Load model
checkpoint = torch.load('pose_transformer.pt')
model = PoseTransformer(**checkpoint['config'])
model.load_state_dict(checkpoint['model_state_dict'])

🔍 Technical Insights​

🧠 Why Joint Tokenization Works​

Traditional Vector Approach:

# Flattened pose vector loses structural information
pose_vector = torch.randn(batch_size, 63)  # 21 joints * 3 dims
output = mlp(pose_vector)  # Limited joint interaction modeling

ViT-Inspired Tokenization:

# Preserve joint structure for better modeling
pose_tokens = pose_vector.view(batch_size, 21, 3)  # Joint tokens
output = transformer(pose_tokens)  # Rich joint interaction modeling

📈 Performance Scaling​

Model Capacity vs. Performance:

  • Embedding Dimension: Linear performance gains up to 256-512
  • Attention Heads: Optimal range 4-16 depending on complexity
  • Layer Depth: Diminishing returns after 6-8 layers for most tasks

🎯 Attention Pattern Analysis​

The transformer learns meaningful attention patterns:

  • Kinematic Chains: Strong attention between connected joints
  • Symmetry: Left-right joint correspondence
  • Coordination: Complex multi-joint movement patterns

This ViT-inspired design represents a breakthrough in pose-to-action translation, achieving unprecedented accuracy through intelligent joint tokenization and transformer-based relationship modeling.