Clarification on use of object geometry in models

Olaronning, the organization behind the 'AI for Industry Challenge', has issued a critical rule clarification for its 2026 competition. In a forum post dated April 16, 2026, they addressed a key technical question from participants: whether AI models can incorporate prior knowledge of an object's physical dimensions. The official answer is yes—competitors are allowed to encode approximate geometry, such as the standard radius, length, and clearance of a plug and socket, directly into their model's design. This clarification removes ambiguity for teams developing perception systems, allowing them to leverage known CAD data or specifications to improve performance, provided this geometric prior is scene-agnostic.

The clarification sets an important boundary: the use of this geometric data must be independent of the specific test scene configuration or any ground-truth pose information provided during the evaluation phase. This ensures the AI is demonstrating genuine understanding and reasoning about objects, not merely memorizing answers from the test set. The ruling is highly relevant for the competition's focus on industrial applications, where robots must reliably manipulate standardized parts. It encourages the development of robust models that blend learned perception with engineered domain knowledge, a hybrid approach often seen in real-world robotics deployments. The forum thread also links to related discussions on tools for 3D asset generation and simulation in platforms like Gazebo and ROS, highlighting the practical ecosystem this challenge operates within.