householdenv 0.0.3

This is an environment for RL purposes

Household-env

This is a gym environment that represents a robot agent in a household environment for RL purposes.

Table of Contents

How to run it
Observation space
Action space
Tasks

How to run it

To run an example of the environment first install it and then run the dummy file.

#!/bin/bash
pip install -e Household-env 
python3 dummy.py

Observation space

The robot has a vision grid of 7x7. The vision grid inputs will return values that represent the content of that cell.

The tasks encoding is binary instead of label encoding (0: task1, 1: task2...) because there is no ordering for the tasks, but the alg. might think there is.

Note: We could do the same for the vision grid, but then the observation space would increase a lot (48*3=147 for for only 7 representable objects )

Num	Observation	Min	Max
0	x_coord_robot	0	19
1	y_coord_robot	0	19
2	task_encoding	0	1
...	...
6	task_encoding	0	1
7	1st action	0	8
8	2nd action	0	8
9	3rd action	0	8
10	4th action	0	8
(vision not yet)
11	vision_grid	0	1
..	..	0	1
58	vision_grid	0	1

Objects will return the following values when within range of the 7x7 vision grid.

Num	Object
0	Nothing
1	wall
2	TV
3	couch
4	bed
5	fridge
6	dishwasher
7	person

Action space

Only one action can be taken at each time step. The Num of the action to be taken is passed as the argument to the step function.

Num	Action	Min	Max
0	move_up	0	1
1	move_down	0	1
2	move_left	0	1
3	move_right	0	1
4	(A) extend_arm	0	1
5	(B) retract_arm	0	1
6	(C) grasp	0	1
7	(D) drop	0	1
8	(E) push	0	1

Tasks

The tasks encoding is binary instead of label encoding (0: task1, 1: task2...) because there is no ordering for the tasks, but the alg. might think there is. We currently use 5 units so up to 31 tasks.

Num	Action	Binary encoded
0	No tasks	00000
1	Turn on TV	00001
2	Bring user a drink	00010
3	Make beds	00011