Anthropomorphic Robotic Arm

“Armed and dangerous”

Overview

This project is still in progress, and I will provide updates as it develops. It began in June 2024, and the current progress reflects about six weeks of work. However, progress has slowed since the semester started, limiting my availability. The design takes inspiration from various sources, but it is entirely my own creation.

The overarching goal for this arm is to create a humanoid arm for embodied machine learning and imitation learning. This leads to the mechanical objective of developing an arm that mimics human size and capabilities while maintaining controllability. Balancing these requirements with low cost is challenging, necessitating careful attention to detail. Just to list a few of the requirements: anthropomorphic morphology, high speed, high torque, backdrivability/torque transparency, precise position and velocity control, and effective torque control.

The current design of the arm encompasses everything up to the wrist joint. Since a human arm has 7 degrees of freedom (DoF) and the wrist adds 2 DoF, this design currently incorporates 5 DoF. The wrist joints and hand will be completed later once this portion of the project is finalized. The hand shown in the pictures is a model I found online as a placeholder. The arm depicted only implements 4 joints. The shoulder consists of 3 rotational joints, but only 2 are active because the joint contained within the body is not currently needed.

Major features

The arm makes ample use of capstan drives. These joints are ideal for their high backdrivability and ease of use. Unlike strainwave gearing or cycloidals, they have excellent backdrivability and are very cheap and do not require machining or expensive components.

One of the goals was to keep all the wires inside the arm. This is possible because capstan actuators can be made as hollow shaft actuators. This is one of the few ways to do this cleanly (or at least without using expensive multi-channel slip rings). The elbow joint is a special type of joint that is more akin to a block and tackle joint. A very similar mechanism is found in the LIMS robot.

The joints make use of moteus r4.11 motor controllers for their excellent field orriented control characteristics. These give position, velocity, and torque control. Combined with the highly backdrivable capstan drives, this setup theoretically gives a great controlability while still remaining low cost. These joints use 8108 KV90 brushless motors for their good torque and performance. Few motors are readily availible in low KV ratings (gimble motors don’t count). The torque can be approximated by 8.3*I_max/KV (source). The controllers can output 11A without thermal management, and 22A with thermal management (source). This means that I can expect about .9Nm without thermals and 1.8Nm with thermals. I won’t go into the details here because I will need to produce a full report on this later, but the best solution I came up with was to use thermal management to maximize torque and minimize my gear ratio. The small gear ratio allows everything to be more backdrivable, and permits the use of low ratio capstan drives.

For manufacturing, this design makes extensive use of 3D printing. Although I know I could have milled some of these parts, many of these would be practically impossible to machine (the design for manufacturing assumed 3D printing). Eventually, I plan to switch from PLA to CF-nylon to increase rigidity and toughness. For the curious, I use a different PLA color every week so that I know which week the prints are from. When I just use black, its difficult to tell if a particular print is the newest one. This leads to a rainbow effect as I switch colors throughout the design.

Just a few CAD sneak peaks. These slices show just how dense this design became. The triple problem of routing electrical power, mechanical power, and airflow became an extremely tight series of constraints to fight.

Results to Date

The arm is absolutely ready to throw hands. This is approximately 17% of maximum speed. Although to be fair I believe that max speed would break the PLA parts until I can upgrade them to CF-nylon.

So far, the arm has proven to be extremely backdrivable and controllable. As an aside, the ‘bounciness’ observed in these few videos is a positive sign, indicating that the actuators are highly backdrivable and exhibit low internal friction. The arm’s full range of capabilities has yet to be tested, and the final design will take time, but this prototype has laid the groundwork. Future work will focus on enhancing the robustness of the existing design and integrating the mechanical system with anticipated machine learning efforts.

This arm combines lessons learned from the large projects I have previously undertaken. It incorporates the electrical knowledge about motor control that I gained while developing the robotic quad actuators and merges it with insights from the rover arm project. The final element I am actively working on right now is the integration of machine learning. I hope to integrate a vision language model with this arm so that I can instruct it to grab specific objects. This project is ongoing, and I will update it occasionally as progress is made.

10/10 would build again, or I guess keep building.