ZXP-S-works / BulletArm

Adding ZXP's envs

BulletArm

• License

BulletArm is a benchmark and learning-environment for robotic manipulation. BulletArm provides a set of standardized benchmark tasks in simulation alongside a collection of baseline algorithms. Please see our paper here.

Table of Contents

1. Requirments
2. Installation
3. Environments
4. Parameters
5. Benchmarks
6. Publications

Requirments

Install

1. Install Python 3.7
2. Clone this repo

   git clone https://github.com/ColinKohler/helping_hands_rl_envs.git
   cd helping_hands_rl_envs
   

3. Install dependencies

   pip install -r requirements.txt 
   

4. Install this package

   pip install .
   

5. Run the block stacking demo

   from bulletarm import env_factory
   # environment parameters
   env_config = {'render': True}
   # create 1 block_stacking environment
   env = env_factory.createEnvs(1, 'block_stacking', env_config)
   env.reset()
   for i in range(5, -1, -1):
       action = env.getNextAction()
       (states_, in_hands_, obs_), rewards, dones = env.step(action)
       input('press enter to continue')
   env.close()

Environments

Open-Loop Environments

• block_picking: The robot needs to pick up all N cubic blocks. The number of blocks N in this environments is configurable.
• block_stacking: (a) The robot needs to stack all N cubic blocks. The number of blocks N in this environments is configurable.
• house_building_1: (b) The robot needs to first stack N-1 cubic blocks then place a triangle block on top of the stack. The number of blocks N in this environments is configurable.
• house_building_2: (c) The robot needs to first place two cubic blocks adjacent to each other, then place a roof on top.
• house_building_3: (d) The robot needs to: 1. place two cubic blocks adjacent to each other; 2. put a cuboid on top of the two cubic blocks; 3. put a roof on top of the cuboid.
• house_building_4: (e) The robot needs to: 1. place two cubic blocks adjacent to each other; 2. put a cuboid on top of the two cubic blocks; 3. put another two cubic blocks on top of the cuboid; 4. put a roof on top of the structure.
• improvise_house_building_2: (f) Similar task as House Building 2, but the fixed cubic blocks are replaced with random shape blocks.
• improvise_house_building_3: (g) Similar task as House Building 3, but the fixed cubic blocks are replaced with random shape blocks.
• block_bin_packing: (h) The robot needs to pack the N blocks in the workspace inside a bin. The number of blocks N in this environments is configurable.
• bottle_tray: (i) The robot needs to arrange six bottles in the tray.
• box_palletizing: (j) The robot needs to palletize N boxes on top of a pallet. The number of boxes N in this environment is configurable (6, 12, or 18). This environments is first proposed in~\cite{transporter}.
• covid_test: (k) The robot needs to supervise three covid tests and gather the test tubes.
• object_grasping: (l) The robot needs to pick up an object in a cluttered scene containing N random objects. The number of objects N in this environment is configurable.

Open-Loop 6D Environments

• ramp_block_stacking: (a) Finish block_stacking in the workspace with two ramps
• ramp_house_building_1: (b) Finish house_building_1 in the workspace with two ramps
• ramp_house_building_2: (c) Finish house_building_2 in the workspace with two ramps
• ramp_house_building_3: (d) Finish house_building_3 in the workspace with two ramps
• ramp_house_building_4: (e) Finish house_building_4 in the workspace with two ramps
• ramp_improvise_house_building_2: (f) Finish improvise_house_building_2 in the workspace with two ramps
• ramp_improvise_house_building_3: (g) Finish improvise_house_building_3 in the workspace with two ramps
• bumpy_house_building_4: (h) Finish house_building_4 in the workspace with a bumpy surface
• bumpy_box_palletizing: (i) Finish box_palletizing in the workspace with a bumpy surface

Close-Loop Environments

• close_loop_block_reaching: The robot needs to place the gripper close to a cubic block.
• close_loop_block_picking: The robot needs to pick up a cubic block.
• close_loop_block_pushing: The robot needs to push the block into a goal area.
• close_loop_block_pulling: The robot needs to pull one of the two blocks to make contact with the other block.
• close_loop_block_in_bowl: The robot needs to pick up a block and place it inside a bowl.
• close_loop_block_stacking: The robot needs to stack N cubic blocks. The number of blocks N in this environment is configurable.
• close_loop_house_building_1: The robot needs to stack N-1 cubic blocks then place a triangle block on top of the stack. The number of blocks N in this environments is configurable.
• close_loop_block_picking_corner: The robot needs to slide the block from the corner and then pick it up.
• close_loop_drawer_opening: The robot needs to pull the handle of the drawer to open it.
• close_loop_clutter_picking: The robot needs to pick up an object in a cluttered scene containing N random objects. The number of objects N in this environment is configurable.

Parameters

Benchmarks

Open-Loop Benchmarks

Prerequisite

1. Install PyTorch (Recommended: pytorch==1.7.0, torchvision==0.8.1)
2. (Optional, required for 6D benchmark) Install CuPy
3. Install other required packages

   pip install -r baseline_requirement.txt
   

4. Goto the baseline directory

   cd helping_hands_rl_envs/helping_hands_rl_baselines/fc_dqn/scripts
   

Open-Loop 3D Benchmark

python main.py --algorithm=[algorithm] --architecture=[architecture] --env=[env]

• Select [algorithm] from: sdqfd (recommended), dqfd, adet, dqn
• Select [architecture] from: equi_asr (recommended), cnn_asr, equi_fcn, cnn_fcn, rot_fcn
• Add --fill_buffer_deconstruct to use deconstruction planner for gathering expert data.

Open-Loop 2D Benchmark

python main.py  --algorithm=[algorithm] --architecture=[architecture] --action_sequence=xyp --random_orientation=f --env=[env]

• Select [algorithm] from: sdqfd (recommended), dqfd, adet, dqn
• Select [architecture] from: equi_fcn (recommended), cnn_fcn
• Add --fill_buffer_deconstruct to use deconstruction planner for gathering expert data.

Open-Loop 6D Benchmark

python main.py  --algorithm=[algorithm] --architecture=[architecture] --action_sequence=xyzrrrp --patch_size=[patch_size] --env=[env]

• Select [algorithm] from: sdqfd (recommended), dqfd, adet, dqn
• Select [architecture] from: equi_deictic_asr (recommended), cnn_asr
• Set [patch_size] to be 40 (required for bumpy_box_palletizing environment) or 24
• Add --fill_buffer_deconstruct to use deconstruction planner for gathering expert data.

Additional Training Arguments

See bulletarm_baselines/fc_dqn/utils/parameters.py

Close-Loop Benchmarks

Prerequisite

1. Install PyTorch (Recommended: pytorch==1.7.0, torchvision==0.8.1)
2. (Optional, required for 6D benchmark) Install CuPy
3. Install other required packages

   pip install -r baseline_requirement.txt
   

4. Goto the baseline directory

   cd helping_hands_rl_envs/helping_hands_rl_baselines/equi_rl/scripts
   

Close-Loop 4D Benchmark

python main.py --algorithm=[algorithm] --env=[env]

• Select [algorithm] from: sac, sacfd, equi_sac, equi_sacfd, ferm_sac, ferm_sacfd, rad_sac, rad_sacfd, drq_sac, drq_sacfd
• To use PER and data augmentation buffer, add --buffer=per_expert_aug

Close-Loop 3D Benchmark

python main.py --algorithm=[algorithm] --action_sequence=pxyz --random_orientation=f --env=[env]

• Select [algorithm] from: sac, sacfd, equi_sac, equi_sacfd, ferm_sac, ferm_sacfd, rad_sac, rad_sacfd, drq_sac, drq_sacfd

Additional Training Arguments

See bulletarm_baselines/equi_rl/utils/parameters.py

Publications

If you like this package and use it in your own work, please cite this repository.