SPARC (Spatial Annotations from Robot Demonstrations at Scale)
SPARC (Spatial Annotations from Robot Demonstrations with Reliability Calibration) is a research-lab category framework and dataset developed by the Intuitive Robots Lab at the Karlsruhe Institute of Technology. It addresses the critical need for high-quality spatial data in robotics by providing an automated auto-labeling pipeline that generates structured annotations like bounding boxes and object trajectories from robot demonstrations. The dataset includes the Interaction-Aware Bench (IA-Bench), which features 1,700 human-annotated demonstrations across 12 diverse robot embodiments, including bimanual setups and Franka platforms. SPARC introduces a unique reliability calibration score to filter noisy labels, significantly improving object localization accuracy compared to standard detection baselines. This open-access resource is designed for researchers developing vision-language-action (VLA) models and grounded robot policies. By leveraging spatio-temporal physical cues, SPARC enables the scaling of training data for complex manipulation tasks in cluttered, real-world environments without requiring exhaustive manual review, thus bridging the gap between raw demonstrations and structured embodied intelligence.
| Year | 2026 |
|---|---|
| Episodes | 1,700 |
| Trajectories | 1,700 |
| Embodiments | Franka, bimanual setups, 12 diverse embodiments |
| Task categories | manipulation, pick-and-place, bimanual, long-horizon |
| License | null |
| Access | open |
| Maintainer | Intuitive Robots Research Team |
| Origin country | DE |
Methodology
SPARC utilizes a risk-aware framework to automatically label robot demonstrations. It isolates manipulated objects using physical cues such as phase-aligned motion and 3D gripper proximity. A spatial filter is applied to suppress the robot's own geometry, ensuring that the annotations focus on the interaction. The core innovation is the Composite Annotation Reliability Score, which captures the likelihood of interaction and serves as a per-annotation estimate of correctness.
Collection
The dataset includes the Interaction-Aware Bench (IA-Bench), a novel benchmark consisting of 1,700 hand-annotated start and target locations of manipulated objects. These demonstrations span 12 diverse robot embodiments and various settings, including bimanual configurations. The data was collected to evaluate model accuracy in grounding the locations of interacted objects within spatio-temporal contexts.
Use Cases
SPARC is intended for training grounded robot policies, embodied foundation models, and vision-language-action (VLA) models. It is particularly effective for motion planning and hierarchical task composition. The reliability scores allow users to trade off annotation quality and dataset scale, making it suitable for training models that must operate in cluttered or visually ambiguous real-world scenes.
Frequently asked questions
What does SPARC stand for?
It stands for Spatial Annotations from Robot Demonstrations with Reliability Calibration.
Who maintains the SPARC dataset?
It is maintained by the Intuitive Robots Lab (IRL) at the Karlsruhe Institute of Technology (KIT).
What is IA-Bench?
IA-Bench is a benchmark within SPARC containing 1,700 human-annotated demonstrations used to measure model accuracy in grounding object locations.
How does SPARC handle noisy labels?
It uses a Composite Annotation Reliability Score based on spatio-temporal physical cues to calibrate the reliability of each automated annotation.
Is the code for SPARC available?
Yes, the code, data, and models are publicly available via the project's GitHub and documentation pages.