During the first three semesters of my master's degree, I was a research assistant at the Neurotechnology Group of ETH Zurich.
We wanted to study the mechanisms behind imitation learning in Marmoset monkeys and needed a
realistically looking robot that could show the monkeys the tasks they were supposed to
learn.
I lead the project to built that robot from conception to deployment, after developing the
specifications and framework, I recruited and supervised students for different parts of the
project.
The robot without cover, demonstrating smooth movements.
Hardware
Marmoset monkeys are very small and have thin arms, therefore the actuators could not be
placed on the robot itself.
Instead, the robot is controlled by strings, which are connected to servo motors in the base
of the robot.
The biggest concern with this kind of design is joints influencing each other, which was
solved by routing the strings for distal
joints through the rotation axis of proximal joints. Where this was not possible, the
strings were routed in a way that both flexor and extensor muscles
were affected equally. To mitigate the change in tension that this causes, the servos are
spring-mounted.
The robot with cover, demonstrating fast movements.
Software
We developed a software framework based on ROS/MoveIt/Gazebo to simulate the robot and plan
its movements. In Rviz, the robot can be controlled by dragging its end effectors. The joint
angles are then generated through inverse kinematics and interpolated between timesteps to
generate smooth movements. When requested through the GUI, the file is sent to the robot and
the movements are executed.