Skip to content

Latest commit

 

History

History
99 lines (67 loc) · 3.43 KB

introduction.md

File metadata and controls

99 lines (67 loc) · 3.43 KB
Raw

Introduction

CTP is a large package containing ROS and deep neural network code for creating and learning task models.

For some more information on the structure of the task planner package, check out the design overview.

Getting started

First follow the installation guide and then try running the simulation on your own. There are two ways of doing this: interactively, through ipython, or via the costar_bullet tool.

You can see the available parameters for costar_bullet with the command:

rosrun costar_bullet start --help

The easiest way to get started using line by line interactive programming is with either ipython or the xonsh shell. You can use the following sample code to bring up the robot with the simple "blocks" task.

import costar_task_plan as ctp
import pybullet as pb

# Create the simulation. Try switching out ur5 for a different robot or blocks
# for a different task.
sim = ctp.simulation.CostarBulletSimulation(robot="ur5", task="blocks", gui=True)

# Start the real-time simulation.
pb.setRealTimeSimulation(True)

# Move the arm around
sim.robot.arm([0,1,0,1,0,1])

# Control the gripper arbitrarily
sim.robot.gripper(0.5)

# Send the gripper to its closed position.
gripper_pos = sim.robot.gripperCloseCommand()
sim.robot.gripper(gripper_pos)

And then interact as you would normally with the PyBullet interface.

Learning

The preferred way of interacting with the sim is through the costar_bullet tool:

rosrun costar_bullet start -h

For example, you can collect a dataset with the command:

rosrun costar_bullet start --robot ur5 --task blocks --agent task --features multi --save -i 10

And then learn a model with:

rosrun costar_bullet start --robot ur5  --features multi --load --model ff_regression

Or the model tool:

rosrun costar_models ctp_model_tool --model ff_regression

Different agents allow you to directly make the robot act in different ways.

Problem Domains

The two primary domains are:

  • Bullet: simulation for various object manipulation tasks, generating task plans, etc.
  • Robotics: ROS version of the above, mostly focusing on the TOM robot.

We have some other domains as well, that do not fully support the CTP library:

  • Road World: drive down a busy road; 2D simulation with second-order dynamics. Not publicly released.
  • Grid World: navigate a busy road in a discrete grid task.
  • Needle Master: steer a needle through a sequence of gates while avoiding obstacles. In many ways this is a simplified driving problem, with an associated set of demonstrations.

More information on the other domains is available.

Bullet

These examples are designed to work with ROS and a simulation of the Universal Robots UR5, KUKA LBR iiwa, or other robot. NOTE THAT THIS FUNCTIONALITY IS STILL IN DEVELOPMENT.

UR5 Simulation

Our examples are based around the costar_bullet package, which uses the open-source Bullet simulator. To start, simply run:

rosrun costar_bullet start

You can run this with the -h or --help flag to get a list of potential arguments. The start command can be configured to bring up a robot and a task. For example, you may want to run:

rosrun costar_bullet start --robot ur5_2_finger --task blocks --gui

To bring up the standard CoSTAR UR5 with Robotiq 85 gripper, a block-stacking task, and a basic Bullet GUI to see things.