Introduces RoboSPA, a large-scale dataset and benchmark for diagnosing VLA models' embodied spatial-procedural reasoning beyond simple task completion.
The query appears to contain a typo; the search context identifies the relevant work as RoboSpatial, not "RoboSPA."
RoboSpatial is a large-scale dataset and benchmark designed to teach spatial understanding to 2D and 3D Vision-Language Models (VLMs) for robotics, addressing the limitation that current VLMs often lack sophisticated spatial reasoning due to training on generic datasets.
Key aspects of RoboSpatial include: Scale and Content: It comprises 1 million images, 5,000 3D scans, and 3 million annotated spatial relationships derived from indoor and tabletop scenes. Core Capabilities: The dataset focuses on three spatial relationships: spatial context (object-space), spatial compatibility (object-object), and spatial configuration (reference frames like ego-, world-, or object-centric). Performance: Models trained on RoboSpatial significantly outperform baselines in downstream tasks such as spatial affordance prediction, spatial relationship prediction, and robot manipulation. Benchmarks: It includes RoboSpatial-Val (a held-out validation subset) and RoboSpatial-Home (manually collected for real-world assessment) to evaluate reasoning in unseen environments.
RoboSPA is presented as a large-scale dataset and benchmark designed to probe whether vision-language-action (VLA) models can perform embodied reasoning in settings more demanding than the simple scenes and short-horizon tasks that dominate current robot-manipulation evaluations. Rather than treating task completion as the primary signal of capability, the work centers on spatial-procedural reasoning: the ability to ground objects and relations in a complex visual scene, infer the procedural structure of a task, and execute a coherent multi-step action sequence under realistic environmental variation.
A key contribution is the reframing of VLA evaluation as a diagnostic problem. RoboSPA appears to stress-test the components that often underlie successful manipulation—spatial grounding, relational reasoning, procedural planning, and long-horizon execution—rather than relying on narrow task templates or highly constrained environments. This matters because many VLA systems can perform well on familiar short-horizon benchmarks while still failing when tasks require integrating scene state, language, and sequential action in more compositional ways.
The broader significance is that RoboSPA provides a more realistic stress test for robot foundation models as they move from demonstration-driven manipulation toward general-purpose embodied agents. If VLA models are to scale beyond controlled laboratory settings, they need to demonstrate not just that they can complete isolated actions, but that they can reason about where things are, what subgoals matter, and how procedures unfold over time. In that sense, RoboSPA is less a single dataset than a benchmarking lens for identifying where current embodied models remain brittle.