---
license: agpl-3.0
task_categories:
- robotics
- video-classification
---

# TIMID: Time-Dependent Mistake Detection in Videos of Robot Executions

[**🌍 Homepage**](https://ropertunizar.github.io/TIMID/) | [**📝 Paper**](https://huggingface.co/papers/2603.09782) | [**💻 Code**](https://github.com/ropertunizar/TIMID)

## Dataset summary
This repository contains the extracted features and model checkpoints associated with [**TIMID: Time-Dependent Mistake Detection in Videos of Robot Executions**](https://ropertunizar.github.io/TIMID/).

**TIMID** is a framework designed to identify and localize temporal mistakes in robotic tasks. Unlike standard action recognition, TIMID focuses on the timing and sequencing of robot executions, detecting when a robot deviates from a "correct" execution path in video streams.

The original raw data is not hosted directly in this repository due to policy reasons. To request access to the original source data, please follow the instructions in the **Accessing the Data** section below.

## Dataset Structure
- **Features**: Pre-computed feature vectors (e.g., embeddings, processed tensors) ready for downstream tasks.
- **Checkpoints**: Model weights saved at various stages of training (found in the `ckpt` folder).

## Usage

### 1. Data preparation
Data and pretrained models are allocated in this repository. You can download them using the Hugging Face CLI:
```bash
hf download nereagallego/TIMID-data --repo-type=dataset --local-dir .
```

### 2. Inference
To run a pre-trained model on one of the datasets (after cloning the [GitHub repo](https://github.com/ropertunizar/TIMID)):
```bash
python main.py --mode infer --model_mode 1 --ckpt_path ckpt/mutex/mutex__7683.pkl --dataset mutex
```
*Datasets: [mutex, ordering, bridge, mutex_real, ordering_real]*

### 3. Training
To train the model on the benchmark:
```bash
python main.py --mode train --model_mode 1 --dataset mutex
```

## Accessing the Data
The original raw dataset is available upon request for research purposes.

To request access:
1. Contact the authors at: **ngallego@unizar.es**
2. Include your name, affiliation, and a brief description of your intended use case.

## Citation

If you find this work, the features, or the checkpoints useful in your research, please cite our work:

```bibtex
@inproceedings{gallego2026timid,
  title={TIMID: Time-Dependent Mistake Detection in Videos of Robot Executions},
  author={Gallego, Nerea and Salanova, Fernando and Mannarano, Claudio and Mahulea, Cristian and Montijano, Eduardo},
  year={2026}
}
```