Eigen RadarAI
Analysis

NVIDIA tests robot grasping with physics-ready simulation assets

NVIDIA has published a worked robot-grasping simulation that assembles an arm, gripper and beverage can from SimReady assets. The exercise checks joints, contact and movement before a scripted transfer in Isaac Sim. It uses a known object position and fixed workspace, with no vision or model of a can being crushed.

Artificial Intelligence··Evening
In an empty robotics test bay, a fixed-base robot arm lifts an unlabelled beverage can in its two-finger gripper and carries it toward a yellow-taped target square on a flat table.

A scripted arm moves a beverage can

NVIDIA published a robot-grasping exercise on November 13 using SimReady, a library of digital assets with physical properties prepared for simulation. Its example combines a robot arm, gripper and beverage can in Isaac Sim, NVIDIA’s robotics simulator. A coding agent helps assemble the scene and generate controls. The programmed sequence brings the gripper towards the can, grasps it, lifts it and moves it to a target before release.[1]

Physical settings are checked before movement

NVIDIA’s single technical walkthrough checks more than a robot’s appearance. Its asset profiles specify mass, collision geometry, joints and the drives controlling those joints. Developers first test small arm movements and opening and closing the gripper, then inspect whether the can rests on the work surface and stays held during transfer. Complete asset folders preserve the components referenced by the scene. Each trial begins from the same reset state so differences in behavior can be compared.[1]

The can starts at a known location

The exercise uses a fixed robot base, clear flat workspace and a can placed at a known position. Motion follows predetermined waypoints without replanning, while contact and friction keep the object in the gripper. Vision is absent and collision checks are limited. The can’s deformation and liquid movement are also outside the simulation. A drive-effort setting chosen in the simulated task is not calibrated to fingertip force. These stated assumptions confine the example to its scripted setup.[1]

References

  1. News sourceNVIDIA / Hugging FaceNVIDIA publishes a robot-grasping experiment using SimReady assets↩1↩2↩3