context_ae_v2_eval_footbag_11Sep_e78b65d9

Context Autoencoder for OOD Detection

This model is a Context Autoencoder trained for out-of-distribution (OOD) detection in robotic tasks. It analyzes temporal patterns in policy activations to detect unusual behavior.

Model Details

• Model Type: Context Autoencoder (LSTM-based temporal autoencoder)
• Training Dataset: [villekuosmanen/drop_footbag_into_dice_tower, villekuosmanen/drop_footbag_into_dice_tower_continuous, villekuosmanen/dAgger_drop_footbag_into_dice_tower_1.0.0, villekuosmanen/dAgger_drop_footbag_into_dice_tower_1.1.0, villekuosmanen/dAgger_drop_footbag_into_dice_tower_1.2.0, villekuosmanen/dAgger_drop_footbag_into_dice_tower_1.3.0, villekuosmanen/dAgger_drop_footbag_into_dice_tower_1.4.0, villekuosmanen/dAgger_drop_footbag_into_dice_tower_1.5.0, villekuosmanen/dAgger_drop_footbag_into_dice_tower_1.6.0, villekuosmanen/eval_footbag_11Sep]
• Layer: model.encoder.layers.3.norm2
• Best Loss: 0.067591

Architecture

• Sequence Length: 100 frames (5.0s at 20 FPS)
• Activation Dimension: 9856
• Hidden Dimension: 4096
• Latent Dimension: 1024
• Number of Layers: 3
• Padding Strategy: repeat_first

Training Configuration

• Learning Rate: 1e-06
• Batch Size: 24
• Epochs: 13
• L2 Penalty: 0.01
• Temporal Smoothness Penalty: 0.001

Usage

Load the model using the ContextAutoencoderWrapper from the codebase:

from missiontracker.checks.context_ae_wrapper import ContextAutoencoderWrapper

# Initialize the wrapper
context_ae = ContextAutoencoderWrapper(
    autoencoder_path="villekuosmanen/context_ae_v2_eval_footbag_11Sep_e78b65d9",  # Will download from Hub
    config_path="villekuosmanen/context_ae_v2_eval_footbag_11Sep_e78b65d9",
    device='cuda',
    max_reconstruction_loss=0.40,  # Adjust threshold as needed
)

# Use for inference
activation_history = [...]  # List of activation tensors
loss, is_anomaly = context_ae.judge_activation(activation_history)

Or directly use with AttentionOODCheck:

from missiontracker.checks.attention_ood import AttentionOODCheck

check = AttentionOODCheck(
    policy=policy,
    context_ae_dir="villekuosmanen/context_ae_v2_eval_footbag_11Sep_e78b65d9",  # Automatically downloads from Hub
    sequence_length=100,
    max_context_ae_loss=0.40,
)

Model Files

• best_model.safetensors: Best model checkpoint (safetensors format)
• config.json: Model configuration and hyperparameters

Citation

If you use this model, please cite:

@misc{context_ae_v2_eval_footbag_11Sep_e78b65d9,
  title={Context Autoencoder for Temporal OOD Detection},
  author={Alpha Robotics},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/villekuosmanen/context_ae_v2_eval_footbag_11Sep_e78b65d9}}
}