Serv-Arm Robotic Manipulator Dataset for Xylo-Aware Spiking Action Selection
Description
This dataset contains joint-space samples from the low-cost Serv-Arm robotic manipulator, organized for the development and evaluation of spiking neural network controllers and neuromorphic action-selection methods. Each sample includes the current configuration of four robotic joints, the desired goal configuration, and the corresponding next joint command. The model input therefore contains eight attributes: four current joint angles and four goal joint angles. For neuromorphic deployment, the original continuous joint-space transitions were converted into nine discrete action classes compatible with the input and output constraints of the Xylo Audio 3 neuromorphic processor: 1. hold 2. base_pos 3. base_neg 4. shoulder_pos 5. shoulder_neg 6. elbow_pos 7. elbow_neg 8. gripper_pos 9. gripper_neg The action label is derived from the dominant normalized joint displacement between the current joint configuration and the next joint command. If all joint displacements are below a predefined threshold, the sample is labeled as hold. Otherwise, the joint with the largest normalized displacement is selected, and the sign of the displacement determines the corresponding positive or negative action. The dataset is divided into training, validation, and test subsets. It includes both the original continuous joint-space representation and a processed representation containing the derived action labels, dominant-joint information, and normalized joint displacements. The processed dataset was used to train and evaluate a compact leaky integrate-and-fire spiking neural network with 8 input channels, 128 hidden neurons, and 9 output classes. The resulting model was evaluated using floating-point simulation, XyloSim, and a physical Xylo Audio 3 neuromorphic processor. The dataset preserves the original class distribution, which is highly imbalanced and reflects the frequency of actions generated from the original joint-space transitions. Researchers may apply class weighting, balanced sampling, or other imbalance-handling strategies during model training while preserving the original validation and test distributions. This dataset supports research in neuromorphic robotics, embedded artificial intelligence, robotic action selection, low-power inference, hardware-aware neural network deployment, simulation-to-hardware evaluation, and reproducible benchmarking of spiking neural networks.
Files
Steps to reproduce
1. Extract the downloaded archive while preserving the directory structure. 2. Use the files in the `original` directory to access the original continuous joint-space data for the training, validation, and test subsets. 3. Each original sample contains the current configuration of the four Serv-Arm joints, the desired goal configuration, and the corresponding next joint command. 4. Use `meta/data_dictionary.csv` to identify the column names, data types, units, and expected value ranges. 5. Use `meta/motor_limits.csv` to access the valid angular ranges adopted for each robotic joint. 6. Run the action-label generation script from the dataset root directory: python scripts/generate_action_labels.py 7. The script converts each continuous transition into one of nine discrete actions by computing the normalized displacement between the current joint configuration and the next joint command. 8. If all joint displacements are below the predefined hold threshold, the sample is labeled as `hold`. Otherwise, the joint with the largest normalized displacement is selected, and the displacement sign determines the corresponding positive or negative action. 9. The generated files are saved in the `processed` directory and contain the original attributes together with the action identifier, action label, dominant-joint information, and normalized displacements. 10. Use `meta/class_mapping.csv` to interpret the numeric identifiers and names of the nine action classes. 11. Preserve the original training, validation, and test subsets when reproducing the experiments reported in the associated study. The original class distribution is intentionally maintained and is highly imbalanced. 12. The processed data can be loaded with standard CSV tools such as pandas: import pandas as pd train = pd.read_csv("processed/train_actions.csv") validation = pd.read_csv("processed/validation_actions.csv") test = pd.read_csv("processed/test_actions.csv") 13. For model reproduction, use the processed files as inputs to a classifier with eight input attributes and nine output classes. The associated study used a leaky integrate-and-fire spiking neural network with 128 hidden neurons. 14. Class weighting or balanced sampling may be applied in new experiments, but the validation and test subsets should remain unchanged for direct comparison with the reported results.
Institutions
- Universidade Federal do AmazonasAmazonas, Manaus