Robot Execution Failures

Original multichannel sensor windows from Robot Execution Failures

Fig. 7 The first non-normal record from each task, with original sample IDs shown above. Force and torque are plotted separately in source units, without normalization. Panel scales differ; sample index is not time. Repeated curves across tasks reflect shared records in the original source.

Source and structure

Official UCI source lists CC BY 4.0. Citation: Lopes, L. & Camarinha-Matos, L. (1998), Robot Execution Failures, UCI Machine Learning Repository, https://doi.org/10.24432/C5M89N.

Each sample has a time-major float32 series of shape (15, 6), zero-based label and zero-based sample_id in original file order. Channels are Fx, Fy, Fz, Tx, Ty, Tz. The source contains integer-valued measurements; the loader validates that conversion to float32 is lossless. No normalization, resampling or windowing is added. The observation window is 315 milliseconds; no sampling-rate assumption is added.

Five configurations retain the source learning problems and their own label spaces:

  • lp1 (default): approach to grasp, 88 records, 4 classes.

  • lp2: transfer of a part, 47 records, 5 classes.

  • lp3: position after transfer failure, 47 records, 4 classes.

  • lp4: approach to ungrasp, 117 records, 3 classes.

  • lp5: motion with part, 164 records, 5 classes.

Each exposes only train. The source has no official train/test split or session/robot identifiers. Do not pool tasks as interchangeable labels. Some identical signal windows occur across tasks; they are preserved. Only the selected text member is read; bundled executables are never run.

Usage

import numpy as np
from stable_datasets.timeseries import RobotExecutionFailures

dataset = RobotExecutionFailures(config_name="lp1", split="train")
signal = np.asarray(dataset[0]["series"], dtype=np.float32)
labels = RobotExecutionFailures.CLASS_NAMES["lp1"]

All configurations share one checksum-verified raw archive, with separate processed caches. Nested sequences may remain lists under formatting; convert explicitly with NumPy or torch.as_tensor.

Citation

@misc{robot_execution_failures_138,
  author = {Lopes, Luis and Camarinha-Matos, Luis},
  title = {Robot Execution Failures},
  year = {1998},
  publisher = {UCI Machine Learning Repository},
  doi = {10.24432/C5M89N},
  url = {https://doi.org/10.24432/C5M89N}
}