Effector Foundation Model Training
A custom effector designed from scratch. End-to-end from data collection to tuning a foundation model and deploying on a Franka robot arm.
Fully 3D-printable UMI effector
This design only requires a camera and rubber bands. Everything else is 3D printed and snap-fits without requiring additional components.
Collecting demonstrations
The effector is guided by hand to collect manipulation data.
Tracking the demonstrations
External markers or SLAM is used to track the relative change in effector pose.
Predicting the trajectory
A foundation model (SmolVLA) is fine-tuned to output the next actions.
Executing on the robot
The predicted motion is executed on a Franka robot.
Handling awkward objects
Not strictly fully 3D printable now, but the composite gripper is soft and easily handles awkward objects.
Better motion tracking
The passive arm tracks the effector's movement through its joints. No SLAM or marker tracking quirks. It just works.
Forward kinematics
The joint measurements reconstruct the end-effector's motion.