yifeidong0 / MTGripperDesign

Multi-task Gripper Co-Design

How to run

For training the RL agent, run the following command:

python3 experiments/train.py

For testing, run the following command:

python3 experiments/test.py

Folder Descriptions

optimizer

Contains the three optimization algorithms:

• grid_search.py: Implements the grid search optimization algorithm.
• mtbo.py: Implements the multi-task Bayesian optimization algorithm.
• random_search.py: Implements the random search optimization algorithm.

sim

Contains the three simulation environments:

push_sim.py (toy example)

• Action Space:
  • Dimensions: 3
  • Meaning: linear and angular velocity change applied to the gripper
• Design Space:
  • Dimensions: 1
  • Meaning: length of the bar-gripper
• Number of Tasks: 2 (pushing a ball, pushing a box)
• Simulator: Custom Pybullet-based simulation

vpush_sim.py

• Action Space:
  • Dimensions: 3
  • Meaning: linear and angular velocity change applied to the gripper
• Design Space:
  • Dimensions: 1
  • Meaning: Angle of the V-shape gripper
• Number of Tasks: 2 (pushing a circle, pushing a polygon)
• Simulator: Custom Box2D-based simulation

ucatch_sim.py

• Action Space:

  • Dimensions: 1
  • Meaning: horizontal linear velocity change applied to the gripper

• Design Space:

  • Dimensions: 5
  • Meaning: Same as action space dimensions
    • d0, d1, d2: Lengths of different parts of the U-shaped gripper
    • alpha0, alpha1: Angles of different parts of the U-shaped gripper

• Number of Tasks: 2 (catching a circle, catching a polygon)

• Simulator: Custom Box2D-based simulation

rl

Contains the visuomotor reinforcement learning modules in the gym environments:

push_env.py

• Environment: Custom Gym environment for the push task.
• Action Space:
  • Dimensions: 3
  • Meaning: linear and angular velocity change applied to the gripper
• Observation Space:
  • Dimensions: Varies (e.g., image data from a camera, state vectors)
  • Meaning: Observations include the current state of the object and possibly visual data
• Reward Structure: Typically designed to encourage the agent to push the object to a target region.

vpush_env.py

• Environment: Custom Gym environment for the variable push task.
• Action Space:
  • Dimensions: 3
  • Meaning: linear and angular velocity change applied to the gripper
• Observation Space:
  • Dimensions: Varies (e.g., image data from a camera, state vectors)
  • Meaning: Observations include the current state of the object and possibly visual data
• Reward Structure: Typically designed to encourage the agent to approach the object (intermediate reward) and push it to a target location with variable initial conditions.

utils

Contains utility functions for the project:

• plot_mtbo.py: Contains the plotting function for Bayesian optimization results.
• plot_mtbo_multi.py: Contains the plotting function for Bayesian optimization (multiple dimensional design space, i.e. >=2) results.

main.py

The main script to run the optimization algorithms. (TODO)

requirements.txt

A list of Python dependencies required to run the project.

Tentative Environments to be Added Soon (TODO)

1. Scooping deformable with a shovel:

   • Simulator: Pybullet
   • Design Space: Spline parameters

2. Wearing masks (linear deformable objects):

   • Simulator: Pybullet
   • Design Space: (To be determined)
   • Observation Space: Key point positions

3. Scooping meat (volumetric deformable objects):

   • Simulator: Pybullet / Isaac Gym
   • Design Space: 3D design parameters