list_rearrange_env.md

List of Rearrange Environments

This document lists many instances of rearrange environments where some of them are designed for training and some of them are designed for testing generalization of a learned policy.

Training Environments

Training environments are classified based on the set of objects and the goal distribution. By default, each training environments are designed to generate a fixed number of objects. If you want to train a policy with a variable number of objects, num_objects should be randomized using the randomization interface.

We recommend to inherit the following default configs for all training environments.

constants: {
    # For the reward, we do the following:
    #  - for every object that gets placed correctly, immediately provide +1
    #  - once all objects are properly placed, the policy needs to hold the goal
    #    state for 2 seconds (in simulation time) until it gets `success_reward` reward
    #  - the next goal is generated
    success_reward: 5.0,
    success_pause_range_s: [0.0, 0.5],
    max_timesteps_per_goal_per_obj: 600,
    vision: True,  # use False if you don't want to use vision observations
    vision_args: {
        image_size: 200,
        camera_names: ['vision_cam_front'],
        mobile_camera_names: ['vision_cam_wrist']
    },
    goal_args: {
        rot_dist_type: 'full',
        randomize_goal_rot: true,
        p_goal_hide_robot:: 1.0,
    },
    # 'obj_pos': l2 distance in meter; 'obj_rot' eular angle distance.
    success_threshold: {'obj_pos': 0.04, 'obj_rot': 0.2},
},
parameters: {
    simulation_params: {
        num_objects: 1,
        max_num_objects: 32,
        object_size: 0.0254,
        used_table_portion: 1.0,
        goal_distance_ratio: 1.0,
        cast_shadows:: false,
        penalty: {
            # Penalty for collisions (wrist camera, table)
            wrist_collision:: 0.0,
            table_collision:: 0.0,
            
            # Penalty for safety stops
            safety_stop_penalty:: 0.0,
        }
    }
}

The table below describes a set of training environment.

Name	File	Config (overwrite)	Description
blocks reach	blocks_reach.py	-	Place end-of-effector of a robot to the target position. This training environment is not compatible with holdout environments.
blocks	blocks_train.py	-	Pushing blocks to targets on the surface of a table
blocks (push + pick-and-place)	blocks_train.py	constants.goal_args.pickup_proba: 0.4	Pushing or pick-and-placing blocks to targets on the surface of a table or in the air
k-composer	composer.py	parameters.simulation_params.num_max_geoms: k	Pushing objects to targets on the surface of a table. Each objects are created by randomly composing [1, k] meshes.
ycb	ycb.py	parameters.simulation_params.mesh_scale: 0.6	Pushing ycb objects to targets on the surface of a table
mixture	mixture.py	constants: {normalize_mesh: True, normalized_mesh_size: 0.05} parameters.simulation_params.mesh_scale: 1.0	Pushing objects to targets on the surface of a table, objects are randomly sample from ycb or simple geom shapes.

Holdout Environments

We design a set of holdout environment to evaluate generalization performance of a learned policy. To make holdout environment compatible with the recommended training environment, we recommend inheriting the same default configs from the training environment.

The table below describes a list of holdout environments and their configs. As a common config for all holdout environments, n-objects holdout environment overwrites parameters.simulation_params.num_objects: n. Configs for some holdout environment can be found in robogym/envs/rearrange/holdouts/configs/. In this case the table will provide a pointer to these configs.

Name	File	Config (overwrite)	Description
n-blocks push	blocks.py	constants.goal_args.rot_dist_type: 'mod90'	push blocks to targets on the surface of the table.
n-blocks flip	blocks.py	constants.goal_args.rot_randomize_type: 'block' parameters.simulation_params.block_appearance: 'openai'	push and flip blocks to targets on the surface of the table. Target block orientation may require flipping the block.
n-blocks duplicate	blocks_duplicate.py	constants.goal_args.rot_dist_type: 'mod90'	push n identical looking blocks to targets on the surface of the table.
n-blocks pick-and-place	blocks_pickandplace.py	constants.goal_args.rot_dist_type: 'mod90'	similar to block push, but one object target is in the air.
n-blocks stack (1 < n < 5)	blocks_stack.py	constants.goal_args.rot_dist_type: 'mod90'	stack n blocks into a tower.
attached blocks	blocks_attached.py	constants.goal_args.rot_dist_type: 'mod90'	rearrange 8 blocks to a particular structured shape (at least one face of each blocks are attached to another).
n-ycb push	ycb.py	-	push ycb objects to targets on the surface of the table.
n-ycb pick-and-place	ycb_pickandplace.py	-	similar to ycb push, but one object target is in the air.
ball_capture	ball_capture.jsonnet	-	see this figure
ball_in_mug	bin_packing.jsonnet	-	see this figure
bin_packing	bin_packing.jsonnet	-	see this figure
bin_unpacking	bin_unpacking.jsonnet	-	see this figure
bookshelf	bookshelf.jsonnet	-	see this figure
chess4	chessboard.py	parameters.simulation_params.num_objects: 4	see this figure
chessboard	chess.jsonnet	-	see this figure
domino5	dominos.py	parameters.simulation_params.num_objects: 5 constants.is_holdout: True	see this figure
dominoes	dominoes.jsonnet	-	see this figure
jenga_cross	jenga/cross.jsonnet	-	see this figure
jenga_tower	jenga/leaning_tower.jsonnet	-	see this figure
jenga_tower_disassemble	jenga/leaning_tower_disassemble.jsonnet	-	see this figure
jenga_tower_stack6	jenga/stack6.jsonnet	-	see this figure
lego_easy_stack2	lego/easy_stack2.jsonnet	-	see this figure
lego_easy_stack3	lego/easy_stack3.jsonnet	-	see this figure
lego_stack2	lego/stack2.jsonnet	-	see this figure
lego_stack2L	lego/stack2L.jsonnet	-	see this figure
lego_stack3	lego/stack3.jsonnet	-	see this figure
lego_stack5	lego/stack5.jsonnet	-	see this figure
rainbow	rainbow.jsonnet	-	see this figure
rainbow_balance	rainbow_build/rainbow_balance.jsonnet	-	see this figure
rainbow_stack2	rainbow_build/rainbow_stack2.jsonnet	-	see this figure
rainbow_stack4	rainbow_build/rainbow_stack4.jsonnet	-	see this figure
rainbow_stack6	rainbow_build/rainbow_stack6.jsonnet	-	see this figure
rainbow_stack_inv2	rainbow_build/rainbow_stack_inv2.jsonnet	-	see this figure
rainbow_stack_inv4	rainbow_build/rainbow_stack_inv4.jsonnet	-	see this figure
rainbow_stack_inv6	rainbow_build/rainbow_stack_inv6.jsonnet	-	see this figure
table_setting	table_setting.py	constants.goal_args.randomize_goal_rot: False constants.success_needed: 1 parameters.simulation_params.num_objects: 5	see this figure
tangram	tangram.jsonnet	-	see this figure
wordblocks	wordblocks.py	constants.goal_args.randomize_goal_rot: False constants.success_needed: 1 parameters.simulation_params.num_objects: 6	see this figure