Basil
Aldoss.

Overview

As part of a University of Iowa senior design team, I am working on a robotic inspection and material-handling application using the Unitree G1 humanoid robot. The objective is to connect visual part inspection with pick-and-place behavior so the robot can direct acceptable and defective parts to their corresponding bins.

The project brings together robot simulation, machine learning, camera-based perception, and physical manipulation. It explores how a humanoid platform can be trained for a practical sorting task and how inspection information can guide the robot's next action.

Simulation and robot learning

We are using NVIDIA Isaac Lab and the GR00T workflow as the foundation for developing pick-and-place behavior. The simulation approach is to represent the robot's workspace, including a table, objects, and destination bins, in a virtual environment that supports training and evaluation before physical robot testing.

The learning workflow is intended to connect task demonstrations, robot policy training, and simulation-based evaluation. Relevant engineering considerations include object placement, robot reach, grasp selection, collision clearance, and repeatable placement in the intended bin.

Camera-based part inspection

A ZED 2i stereo camera mounted on the robot provides the visual input for the inspection portion of the project. We are developing a machine-learning workflow to distinguish acceptable parts from defective parts and use that classification to select the appropriate sorting destination.

The inspection and manipulation functions need to work together: identifying a part's category is useful only if that decision can be connected to a reliable grasp and placement action.

Target system workflow

1. Observe the partAcquire visual input from the robot-mounted ZED 2i camera.
2. Classify its conditionUse a machine-learning inspection model to distinguish acceptable and defective parts.
3. Select the destinationMap the inspection decision to the corresponding bin.
4. Pick and placeUse the G1's trained manipulation behavior to grasp the part and place it in the selected bin.
5. Evaluate the taskAssess whether the part was classified correctly and placed successfully.

Engineering focus

• Representing a physical sorting task in a simulation environment for robot learning.
• Connecting part inspection decisions with robotic material-handling actions.
• Considering workspace geometry, object accessibility, grasping, and bin placement.
• Evaluating perception and manipulation together as one integrated system.

Tools and hardware

Unitree G1 humanoid robot · NVIDIA Isaac Lab · GR00T workflow · ZED 2i stereo camera · Computer vision · Machine learning