ArmstrongUMI
A data capture device with a compact handle design, enabling human operators to capture manipulation data in spaces accessible to robotic arms. For additional details, see our hardware-focused report.
Responsibilities
-
Redesigned the handle and actuation mechanism of a UMI-based gripper tailored for our application
-
Collaborated with software and hardware teams to implement required hardware features
-
Support the software team in collecting manipulation data and evaluating robotic performance
The original UMI gripper was a strong reference, but its bulky handle and wide top plate limited access in dishwashing environments, where reaching deep into bus tubs and sink basins is essential, even when the robot arm itself could reach the object.
So we redesigned the gripper to match the actual robot end-effector geometry, ensuring that any grasp the robot could perform would also be feasible during data collection with our gripper. This led to a thumb-driven pinching interaction instead of a vertical handle.
.jpg)
For localization, we are using a Quest controller. However, we plan to replace this with ArUco marker-based tracking to reduce both weight and cost. In its current configuration, the unit weighs about 450g and costs around $200 per unit.
One of the most striking aspects of this project was adopting a new AI-driven workflow. To communicate a new direction to investors, I compressed ideation, design, and video production into a two-week sprint, something that would not have been possible without it.
Currently, the software team is using the device I designed to collect real-world manipulation episodes for policy learning. For additional details, see our hardware-focused report.
Further process and information available upon request.