Deep Learning Model for Forward Dynamics of Robot Arm
I developed a deep learning model to learn the forward dynamics of a 3-link robotic arm to imitate the behavior of another arm.
Problem Statement
There is a 3-link arm (arm_teacher) operates with a provided ground truth forward dynamics model, my objective was to train another arm to replicate this behavior accurately.
My Approach
Training Data Collection
The first part of the project involves collecting a dataset that will be used to train the deep learning model. This dataset consists of state-action pairs and their resulting new states, generated using the ground truth arm. The comprehensive dataset is 250,000 state-action pairs.
State: (6,1) dimensional Numpy array
3 joint positions in radians + 3 joint velocities in rad/second
Action: (3,1) dimensional Numpy array 3 toques (in Nm) applied to the three joints respectively Problem is simplified by just having torque applied on the first joint.
More details:
Learning the forward dynamics
In this part, I trained a DNN based on the data collected from the arm_teacher to learn the forward dynamics of the arm.
More details:
Difficulties
Why the output of the network is acceleration rather than the state?
My justification:
Result
Evaluated the model on 3 different types of tests, 5 tests of each type. Achieved an average MSE of 1.87-e5.
More details of the test can be found in the project video!