The GummiArm Project: A Replicable and Variable-Stiffness Robot Arm for Experiments on Embodied AI

Area of Science:

• Robotics
• Artificial Intelligence
• Embodied AI

Background:

• Research robots for Embodied AI require physical exploration and learning from failure.
• Existing robots are often too stiff, expensive, and fragile for safe, repeated experimentation.

Purpose of the Study:

• To develop an open-source, affordable, and robust soft robot arm for Embodied AI research.
• To enable easier replication of robotics experiments and concurrent design of AI 'brain' and 'body'.

Main Methods:

• Designed a 6-year project resulting in the GummiArm, featuring human-inspired tendon actuation.
• Incorporated variable-stiffness and damped actuation for safety and new research avenues.
• Utilized hobby-grade 3D printing for arm structure and stretchable composite tendons for impact resistance.

Main Results:

• The GummiArm is low-cost (<$6000), with easily replaceable 3D-printed parts (<$20 filament).
• The arm exhibits dexterity for simple manipulation and is robust to impacts.
• Demonstrated a repair cycle of minutes for damaged components.

Conclusions:

• The GummiArm facilitates safer physical exploration and learning in Embodied AI research.
• Its affordability, durability, and ease of replication promote productivity and reproducibility.
• Enables concurrent development of robotic bodies and AI control systems.