AI xPress

AI and computer science research.

Back to Foragentis

Robotics & Embodied AI

FoundationGrasp outperformed earlier methods on novel objects and tasks in robot grasping tests

Robotics & Embodied AI November 4, 2026

A humanoid robot in a lab among tools

In a 2025 paper in IEEE Transactions on Automation Science and Engineering, researchers from Southern University of Science and Technology and Georgia Institute of Technology describe a system called FoundationGrasp. The system draws on open-ended knowledge from foundation models to choose grasps that fit a later manipulation task. In the study's experiments, FoundationGrasp outperformed existing methods on novel object instances, object classes, and tasks outside its training set. The team also validated the system in real-robot grasping and manipulation tests on a 7-DoF robotic arm, which has seven degrees of freedom.

What is task-oriented grasping?

Task-oriented grasping means choosing a grip that is compatible with the next task, not just any grip that holds the object. For example, a robot may need to grasp kitchenware in a way that lets it clean the item, or grasp a part in a way that lets it assemble something. FoundationGrasp is designed to learn this skill for objects and tasks it has not seen.

How did the researchers test FoundationGrasp?

The benchmark experiments used the authors' contributed Language and Vision Augmented TaskGrasp dataset, called LaViA-TaskGrasp. On that dataset, FoundationGrasp outperformed existing methods when generalizing to novel object instances, object classes, and tasks out of the training set. The authors also ran real-robot grasping and manipulation experiments on a 7-DoF robotic arm.

What do the findings suggest?

The authors write that solving generalizable task-oriented grasping could significantly improve a robot's level of automation and intelligence in tool manipulation for household and industrial tasks. They give examples such as cleaning kitchenware and assembling parts in industrial contexts. The benchmark and robot results suggest that foundation-model knowledge gives the system a way to handle unseen objects and tasks.

FAQ

What is a foundation model?

A foundation model is an AI model trained on broad data, so it carries general knowledge. In this paper, FoundationGrasp uses that open-ended knowledge to learn grasping skills that are not limited to a fixed set of objects or tasks.

What does "out of the training set" mean?

It means the objects or tasks were not part of the data used to train the system. The paper tested FoundationGrasp on novel object instances, object classes, and tasks of that kind.

Where was the system tested?

The benchmark tests used the authors' LaViA-TaskGrasp dataset, and the real-robot tests used a 7-DoF robotic arm.

Limits and what's next

This is one study. The benchmark results came from the authors' contributed LaViA-TaskGrasp dataset, and the real-robot tests used a 7-DoF robotic arm. The abstract does not report an independent replication by another team. The authors state that the code, data, appendix, and video are publicly available, which lets other researchers inspect the work.

Source: Chao Tang et al., “FoundationGrasp: Generalizable Task-Oriented Grasping With Foundation Models,” IEEE Transactions on Automation Science and Engineering, 2025.