v0.1.0-rc12 Improved Goal Tracking + Table Cleaning Tasks w/ Fixed object ρ₀

1. now sample goal offsets for tracking targets once per episode (upon reset, saved in goal_offsets cache)
2. add the following sub-tasks in the table cleaning taskset for faster experiments

# Fixing the objects in-place
register(id='Clean-i-fixed-v0', entry_point=vec_clean_env, kwargs=dict(
    initial_qpos={'object1:joint': [1.35, 0.75, .45, 0, 0., 0., 0.]},
    obj_reset={'object0': dict(pos=[1.35, 0.7, 0.42478468]),
               'object1': dict(pos=[1.35, 0.8, 0.42478468]), },
    goal_sampling={'object0': dict(target="box", range=0, ),
                   'object1': dict(target="object1", range=0, offset=[0, 0, 0]), },
), **kw)

register(id='Clean-ii-fixed-v0', entry_point=vec_clean_env, kwargs=dict(
    obj_reset={'object0': dict(track='box', avoid=['gripper'], range=0, h=0.43456914),
               'object1': dict(pos=[1.35, 0.8, 0.42478468]), },
    goal_sampling={'object0': dict(target="box", range=0),
                   'object1': dict(target="object0", range=0), },
), **kw)

register(id='Clean-train-fixed-v0', entry_point=SampleEnv,
         kwargs={'fetch:Clean-i-fixed-v0': 0.5, 'fetch:Clean-ii-fixed-v0': 0.5, }, **kw)

register(id='Clean-train-fixed-80-20-v0', entry_point=SampleEnv,
         kwargs={'fetch:Clean-i-fixed-v0': 0.8, 'fetch:Clean-ii-fixed-v0': 0.2, }, **kw)

register(id='Clean-fixed-v0', entry_point=vec_clean_env, kwargs=dict(
    obj_reset={'object0': dict(pos=[1.35, 0.7, 0.42478468]),
               'object1': dict(pos=[1.35, 0.8, 0.42478468]), },
    goal_sampling={'object0': dict(target="box", range=0),
                   'object1': dict(target="box", range=0), },
), **kw)

v0.1.0-rc11 Improve Clean task initial condition and Goal height

Task: Cleaning Table

Primitives

1. place A into box, both on table.
2. place B into box, with A already inside the box.

Training Mixture: 1 + 2, 50:50.

Test: Both A and B on table, and place both into the box. (edited)

Name	Observation Spec	Goal Init/Comment
Clean-i-v0	achieved_goal: (6,) desired_goal: (6,) observation: (43,)	place the first object into the box
Clean-ii-v0	achieved_goal: (6,) desired_goal: (6,) observation: (43,)	place the second object into the box while the first one is already in there.
Clean-train-v0	achieved_goal: (6,) desired_goal: (6,) observation: (43,)	50/50 mixture of both
Clean-v0	achieved_goal: (6,) desired_goal: (6,) observation: (43,)	Start with both on the table, place both into the box.

twin_box_envs = [
    'fetch:Clean-i-v0',
    'fetch:Clean-ii-v0',
    'fetch:Clean-train-v0',
    'fetch:Clean-v0',
    'fetch:Clean-aside-i-v0',
    'fetch:Clean-aside-ii-v0',
    'fetch:Clean-aside-train-v0',
    'fetch:Clean-aside-v0',
]

Clean-i-v0	distribution
Clean-ii-v0	distribution
Clean-train-v0	distribution
Clean-v0	distribution
Clean-aside-i-v0	distribution
Clean-aside-ii-v0	distribution
Clean-aside-train-v0	distribution
Clean-aside-v0	distribution

v0.1.0-rc9 Fix returned goals for targets that depends on other objects

It turned out that goals that depend on the position of other objects are wrong right after env.reset() call, but they are good otherwise (for the rest of the trajectory).

v0.1.0-rc8 Remove `ml_logger` from dependencies

See rc7 for env list and installation guides.

v0.1.0-rc7 Add Updated Box Single Taskset

write_protected: true

Cloning and Installing A Specific Version

For those who do not use pip, you can pull/checkout specific versions of the repository using a git tag. Github integrates with git, such that it shows tagged versions under the <project-repo>/releases URL. You can check out the releases for this repo at https://github.com/geyang/gym-fetch/releases. To pull a specific version, just run:

git checkout v0.1.0-rc5

After git clone, the structure of the project is

$ tree . -L 1
gym-fetch
├── fetch
└── specs

To install, you can use pip

pip install -e .

Alternatively, you can include the ./gym-fetch directory as part of your PYTHONPATH.

Box Single Task Taskset

Simple Goal Distributions

The goal distribution of these tasks are pure. Tasks with place postfix have the goal
initialized at the center of the box, whereas the other two uses the goal distribution
from standard gym FetchPickAndPlace-v1 environment.

Name	Observation Spec	Goal Init/Comment
Box-fixed-v0	achieved_goal: (3,) desired_goal: (3,) observation: (25,)	box is in the middle
Box-aside-v0	achieved_goal: (3,) desired_goal: (3,) observation: (25,)	box is on the side
Box-fixed-place-v0	achieved_goal: (3,) desired_goal: (3,) observation: (25,)	box is on the side
Box-aside-place-v0	achieved_goal: (3,) desired_goal: (3,) observation: (25,)	box is on the side

Mixture Goal Distributions

These two training environments uses a mixture of 20/80% of pick-and-place distribution vs target inside the box.

Name	Observation Spec	Goal Init/Comment
Box-fixed-place-train-v0	achieved_goal: (3,) desired_goal: (3,) observation: (25,)	box is on the side
Box-aside-place-train-v0	achieved_goal: (3,) desired_goal: (3,) observation: (25,)	box is on the side

Details of Each Task

box_envs = ['fetch:Box-aside-v0',
            'fetch:Box-aside-place-v0',
            'fetch:Box-aside-place-train-v0',
            'fetch:Box-fixed-v0',
            'fetch:Box-fixed-place-v0',
            'fetch:Box-fixed-place-train-v0']

Box-aside-v0	distribution
Box-aside-place-v0	distribution
Box-aside-place-train-v0	distribution
Box-fixed-v0	distribution
Box-fixed-place-v0	distribution
Box-fixed-place-train-v0	distribution

Adding StackTwo Task

This release includes the new StackTwo environment, with a new ρ₀ that makes it easy to learn.

To install:

pip install gym-fetch==0.1.0rc2

Or you can install from github (might be slower b/c images)

pip install git+git://github.com/geyang/gym-fetch.git#v0.1.0-rc2

Adding Single Task Environments

This is the first public release including all of the single task environments.