calebescobedo/sensor-diffusion-policy-table-camera-epoch220

Sensor Diffusion Policy - Epoch 220

Diffusion policy model trained on proximity sensor data with table camera images.

Model Details

• Model Type: Diffusion Policy
• Training Epochs: 220/300
• Horizon: 16 steps
• Observation Steps: 1
• Action Steps: 8

Inputs

The model expects the following inputs:

1. observation.state (STATE)

• Shape: (batch, 1, 7)
• Description: Joint positions for 7-DOF arm
• Normalization: Min-max normalized using dataset statistics

2. observation.goal (STATE)

• Shape: (batch, 1, 3)
• Description: Goal cartesian position (X, Y, Z in meters)
• Normalization: Min-max normalized using dataset statistics

3. observation.images.table_camera (VISUAL)

• Shape: (batch, 1, 3, 480, 640)
• Description: RGB images from table camera
• Normalization: Mean-std normalized (normalized to [0, 1] then mean-std)

4. observation.proximity (STATE)

• Shape: (batch, 1, 128)
• Description: Encoded proximity sensor latent (37 sensors × 8×8 depth maps → 128-dim)
• Normalization: Min-max normalized using dataset statistics
• Note: Requires pretrained ProximityAutoencoder encoder

Outputs

action (ACTION)

• Shape: (batch, 7)
• Description: Joint positions (7-DOF) for the next timestep
• Type: Joint positions (not velocities) - these are the next positions the robot should move to
• Normalization: Output is unnormalized (raw joint positions in radians)

Usage

from lerobot.policies.diffusion.modeling_diffusion import DiffusionPolicy
from lerobot.policies.factory import make_pre_post_processors
import torch

# Load model and processors
repo_id = "calebescobedo/sensor-diffusion-policy-table-camera-epoch220"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

policy = DiffusionPolicy.from_pretrained(repo_id)
policy.eval()
policy.to(device)

preprocessor, postprocessor = make_pre_post_processors(
    policy_cfg=policy.config,
    pretrained_path=repo_id
)

# Prepare input batch
# Note: You need to encode proximity sensors using the ProximityAutoencoder first
batch = {
    "observation.state": torch.tensor([...]),  # Shape: (batch, 1, 7)
    "observation.goal": torch.tensor([...]),    # Shape: (batch, 1, 3)
    "observation.images.table_camera": torch.tensor([...]),  # Shape: (batch, 1, 3, 480, 640)
    "observation.proximity": torch.tensor([...]),  # Shape: (batch, 1, 128) - encoded
}

# Predict action
with torch.no_grad():
    batch_processed = preprocessor(batch)  # Normalizes inputs
    actions = policy.select_action(batch_processed)  # Returns normalized actions
    actions = postprocessor(actions)  # Unnormalizes to raw joint positions

# actions shape: (batch, 7) - joint positions in radians

Proximity Sensor Encoding

The proximity sensors must be encoded before use. You need to load the ProximityAutoencoder:

from architectures.proximity_autoencoder import ProximityAutoencoder
import torch

# Load proximity encoder
encoder_path = "path/to/proximity_autoencoder.pth"
ae_model = ProximityAutoencoder(num_sensors=37, depth_channels=1, latent_dim=128, use_attention=True)
ae_model.load_state_dict(torch.load(encoder_path, map_location='cpu'))
proximity_encoder = ae_model.encoder
proximity_encoder.eval()

# Encode proximity sensors (37 sensors × 8×8 depth maps)
# raw_proximity shape: (batch, 37, 8, 8)
encoded_proximity = proximity_encoder(raw_proximity)  # Shape: (batch, 128)

Dataset Statistics

Dataset statistics are included in config.json under the dataset_stats key. These are used for normalization/unnormalization and were computed from the training dataset:

• /home/caleb/datasets/sensor/roboset_20260117_014645/*.h5 (20 files, ~500 trajectories)

Training Details

• Dataset: Sensor dataset with proximity sensors and table camera
• Training Loss: ~0.003-0.004 (at epoch 220)
• Optimizer: Adam (LeRobot preset)
• Learning Rate: From LeRobot optimizer preset
• Mixed Precision: Enabled (AMP)
• Data Augmentation: State noise (30% prob, scale=0.005), Action noise (30% prob, scale=0.0005), Goal noise (30% prob)

Model Architecture

• Vision Backbone: ResNet18
• Diffusion Steps: 100
• Noise Scheduler: DDPM with squaredcos_cap_v2 beta schedule
• Total Parameters: ~261M

Citation

If you use this model, please cite:

@misc{sensor-diffusion-policy-epoch220,
  author = {Caleb Escobedo},
  title = {Sensor Diffusion Policy - Epoch 220},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/calebescobedo/sensor-diffusion-policy-table-camera-epoch220}}
}