Safe Multi-Agent Isaac Gym benchmark is for safe multi-agent reinforcement learning research.
The README is organized as follows:
- About this repository
- Installation
- Running the benchmarks
- Select an algorithm
- Select tasks
This repository is an extention of DexterousHands which is from PKU MARL research team, Safe MAIG contains complex dexterous hand RL environments for the NVIDIA Isaac Gym high performance environments described in the NeurIPS 2021 Datasets and Benchmarks paper.
🌟 This repository is under actively development. We appreciate any constructive comments and suggestions. If you have any questions, please feel free to email <gshangd[AT]foxmail.com>.
Details regarding installation of IsaacGym can be found here. We currently support the Preview Release 3
version of IsaacGym.
The code has been tested on Ubuntu 18.04 with Python 3.7. The minimum recommended NVIDIA driver
version for Linux is 470
(dictated by support of IsaacGym).
It uses Anaconda to create virtual environments. To install Anaconda, follow instructions here.
Ensure that Isaac Gym works on your system by running one of the examples from the python/examples
directory, like joint_monkey.py
. Follow troubleshooting steps described in the Isaac Gym Preview 2
install instructions if you have any trouble running the samples.
Once Isaac Gym is installed and samples work within your current python environment, install this repo:
pip install -e .
To train your first policy, run this line:
python train.py --task=ShadowHandOver --algo=macpo
Or run the lines:
chmod +x ./run_experiments.sh
./run_experiments.sh
If you want to turn off the video, set headless as True, e.g.,
python train.py --task=ShadowHandOver --algo=macpo --headless=True
To select an algorithm, pass --algo=ppo/mappo/happo/hatrpo
as an argument:
python train.py --task=ShadowHandOver --algo=macpo
At present, we only support these four algorithms.
Source code for tasks can be found in dexteroushandenvs/tasks
.
Until now we only suppose the following environments:
Environments | ShadowHandOver | ShadowHandCatchUnderarm | ShadowHandTwoCatchUnderarm | ShadowHandCatchAbreast | ShadowHandOver2Underarm |
---|---|---|---|---|---|
Description | These environments involve two fixed-position hands. The hand which starts with the object must find a way to hand it over to the second hand. | These environments again have two hands, however now they have some additional degrees of freedom that allows them to translate/rotate their centre of masses within some constrained region. | These environments involve coordination between the two hands so as to throw the two objects between hands (i.e. swapping them). | This environment is similar to ShadowHandCatchUnderarm, the difference is that the two hands are changed from relative to side-by-side posture. | This environment is is made up of half ShadowHandCatchUnderarm and half ShadowHandCatchOverarm, the object needs to be thrown from the vertical hand to the palm-up hand |
Actions Type | Continuous | Continuous | Continuous | Continuous | Continuous |
Total Action Num | 40 | 52 | 52 | 52 | 52 |
Action Values | [-1, 1] | [-1, 1] | [-1, 1] | [-1, 1] | [-1, 1] |
Action Index and Description | detail | detail | detail | detail | detail |
Observation Shape | (num_envs, 2, 211) | (num_envs, 2, 217) | (num_envs, 2, 217) | (num_envs, 2, 217) | (num_envs, 2, 217) |
Observation Values | [-5, 5] | [-5, 5] | [-5, 5] | [-5, 5] | [-5, 5] |
Observation Index and Description | detail | detail | detail | detail | detail |
State Shape | (num_envs, 2, 398) | (num_envs, 2, 422) | (num_envs, 2, 422) | (num_envs, 2, 422) | (num_envs, 2, 422) |
State Values | [-5, 5] | [-5, 5] | [-5, 5] | [-5, 5] | [-5, 5] |
Rewards | Rewards is the pose distance between object and goal. You can check out the details here | Rewards is the pose distance between object and goal. You can check out the details here | Rewards is the pose distance between object and goal. You can check out the details here | Rewards is the pose distance between two object and two goal, this means that both objects have to be thrown in order to be swapped over. You can check out the details here | Rewards is the pose distance between object and goal. You can check out the details here |
Demo |
These environments involve two fixed-position hands. The hand which starts with the object must find a way to hand it over to the second hand. To use the HandOver environment, pass --task=ShadowHandOver
Index | Description |
---|---|
0 - 23 | shadow hand dof position |
24 - 47 | shadow hand dof velocity |
48 - 71 | shadow hand dof force |
72 - 136 | shadow hand fingertip pose, linear velocity, angle velocity (5 x 13) |
137 - 166 | shadow hand fingertip force, torque (5 x 6) |
167 - 186 | actions |
187 - 193 | object pose |
194 - 196 | object linear velocity |
197 - 199 | object angle velocity |
200 - 206 | goal pose |
207 - 210 | goal rot - object rot |
The shadow hand has 24 joints, 20 actual drive joints and 4 underdrive joints. So our Action is the joint Angle value of the 20 dimensional actuated joint.
Index | Description |
---|---|
0 - 19 | shadow hand actuated joint |
Rewards is the pose distance between object and goal, and the specific formula is as follows:
goal_dist = torch.norm(target_pos - object_pos, p=2, dim=-1)
quat_diff = quat_mul(object_rot, quat_conjugate(target_rot))
rot_dist = 2.0 * torch.asin(torch.clamp(torch.norm(quat_diff[:, 0:3], p=2, dim=-1), max=1.0))
dist_rew = goal_dist
reward = torch.exp(-0.2*(dist_rew * dist_reward_scale + rot_dist))
These environments again have two hands, however now they have some additional degrees of freedom that allows them to translate/rotate their centre of masses within some constrained region. To use the HandCatchUnderarm environment, pass --task=ShadowHandCatchUnderarm
Index | Description |
---|---|
0 - 23 | shadow hand dof position |
24 - 47 | shadow hand dof velocity |
48 - 71 | shadow hand dof force |
72 - 136 | shadow hand fingertip pose, linear velocity, angle velocity (5 x 13) |
137 - 166 | shadow hand fingertip force, torque (5 x 6) |
167 - 192 | actions |
193 - 195 | shadow hand transition |
196 - 198 | shadow hand orientation |
199 - 205 | object pose |
206 - 208 | object linear velocity |
209 - 211 | object angle velocity |
212 - 218 | goal pose |
219 - 222 | goal rot - object rot |
Similar to the HandOver environments, except now the bases are not fixed and have translational and rotational degrees of freedom that allow them to move within some range.
Index | Description |
---|---|
0 - 19 | shadow hand actuated joint |
20 - 22 | shadow hand actor translation |
23 - 25 | shadow hand actor rotation |
Rewards is the pose distance between object and goal, and the specific formula is as follows:
goal_dist = torch.norm(target_pos - object_pos, p=2, dim=-1)
quat_diff = quat_mul(object_rot, quat_conjugate(target_rot))
rot_dist = 2.0 * torch.asin(torch.clamp(torch.norm(quat_diff[:, 0:3], p=2, dim=-1), max=1.0))
dist_rew = goal_dist
reward = torch.exp(-0.2*(dist_rew * dist_reward_scale + rot_dist))
This environment is is made up of half ShadowHandCatchUnderarm and half ShadowHandCatchOverarm, the object needs to be thrown from the vertical hand to the palm-up hand. To use the HandCatchUnderarm environment, pass --task=ShadowHandCatchOver2Underarm
Index | Description |
---|---|
0 - 23 | shadow hand dof position |
24 - 47 | shadow hand dof velocity |
48 - 71 | shadow hand dof force |
72 - 136 | shadow hand fingertip pose, linear velocity, angle velocity (5 x 13) |
137 - 166 | shadow hand fingertip force, torque (5 x 6) |
167 - 192 | actions |
193 - 195 | shadow hand transition |
196 - 198 | shadow hand orientation |
199 - 205 | object pose |
206 - 208 | object linear velocity |
209 - 211 | object angle velocity |
212 - 218 | goal pose |
219 - 222 | goal rot - object rot |
Similar to the HandOver environments, except now the bases are not fixed and have translational and rotational degrees of freedom that allow them to move within some range.
Index | Description |
---|---|
0 - 19 | shadow hand actuated joint |
20 - 22 | shadow hand actor translation |
23 - 25 | shadow hand actor rotation |
Rewards is the pose distance between object and goal, and the specific formula is as follows:
goal_dist = torch.norm(target_pos - object_pos, p=2, dim=-1)
# Orientation alignment for the cube in hand and goal cube
quat_diff = quat_mul(object_rot, quat_conjugate(target_rot)
reward = (0.3 - goal_dist - quat_diff)
These environments involve coordination between the two hands so as to throw the two objects between hands (i.e. swapping them). This is necessary since each object's goal can only be reached by the other hand. To use the HandCatchUnderarm environment, pass --task=ShadowHandTwoCatchUnderarm
Index | Description |
---|---|
0 - 23 | shadow hand dof position |
24 - 47 | shadow hand dof velocity |
48 - 71 | shadow hand dof force |
72 - 136 | shadow hand fingertip pose, linear velocity, angle velocity (5 x 13) |
137 - 166 | shadow hand fingertip force, torque (5 x 6) |
167 - 192 | actions |
193 - 195 | shadow hand transition |
196 - 198 | shadow hand orientation |
199 - 205 | object1 pose |
206 - 208 | object1 linear velocity |
210 - 212 | object1 angle velocity |
213 - 219 | goal1 pose |
220 - 223 | goal1 rot - object1 rot |
224 - 230 | object2 pose |
231 - 233 | object2 linear velocity |
234 - 236 | object2 angle velocity |
237 - 243 | goal2 pose |
244 - 247 | goal2 rot - object2 rot |
Similar to the HandOver environments, except now the bases are not fixed and have translational and rotational degrees of freedom that allow them to move within some range.
Index | Description |
---|---|
0 - 19 | shadow hand actuated joint |
20 - 22 | shadow hand actor translation |
23 - 25 | shadow hand actor rotation |
Rewards is the pose distance between two object and two goal, this means that both objects have to be thrown in order to be swapped over. The specific formula is as follows:
goal_dist = torch.norm(target_pos - object_pos, p=2, dim=-1)
goal_another_dist = torch.norm(target_another_pos - object_another_pos, p=2, dim=-1)
# Orientation alignment for the cube in hand and goal cube
quat_diff = quat_mul(object_rot, quat_conjugate(target_rot))
rot_dist = 2.0 * torch.asin(torch.clamp(torch.norm(quat_diff[:, 0:3], p=2, dim=-1), max=1.0))
quat_another_diff = quat_mul(object_another_rot, quat_conjugate(target_another_rot))
rot_another_dist = 2.0 * torch.asin(torch.clamp(torch.norm(quat_another_diff[:, 0:3], p=2, dim=-1), max=1.0))
dist_rew = goal_dist
reward = torch.exp(-0.2*(dist_rew * dist_reward_scale + rot_dist)) + torch.exp(-0.2*(goal_another_dist * dist_reward_scale + rot_another_dist))
This environment is similar to ShadowHandCatchUnderarm, the difference is that the two hands are changed from relative to side-by-side posture.. To use the HandCatchAbreast environment, pass --task=ShadowHandCatchAbreast
Index | Description |
---|---|
0 - 23 | shadow hand dof position |
24 - 47 | shadow hand dof velocity |
48 - 71 | shadow hand dof force |
72 - 136 | shadow hand fingertip pose, linear velocity, angle velocity (5 x 13) |
137 - 166 | shadow hand fingertip force, torque (5 x 6) |
167 - 192 | actions |
193 - 195 | shadow hand transition |
196 - 198 | shadow hand orientation |
199 - 205 | object pose |
206 - 208 | object linear velocity |
209 - 211 | object angle velocity |
212 - 218 | goal pose |
219 - 222 | goal rot - object rot |
Similar to the HandOver environments, except now the bases are not fixed and have translational and rotational degrees of freedom that allow them to move within some range.
Index | Description |
---|---|
0 - 19 | shadow hand actuated joint |
20 - 22 | shadow hand actor translation |
23 - 25 | shadow hand actor rotation |
Rewards is the pose distance between object and goal, and the specific formula is as follows:
goal_dist = torch.norm(target_pos - object_pos, p=2, dim=-1)
quat_diff = quat_mul(object_rot, quat_conjugate(target_rot))
rot_dist = 2.0 * torch.asin(torch.clamp(torch.norm(quat_diff[:, 0:3], p=2, dim=-1), max=1.0))
dist_rew = goal_dist
reward = torch.exp(-0.2*(dist_rew * dist_reward_scale + rot_dist))
If you find the repository useful, please cite the paper:
@article{gu2023safe,
title={Safe Multi-Agent Reinforcement Learning for Multi-Robot Control},
author={Gu, Shangding and Kuba, Jakub Grudzien and Chen, Yuanpei and Du, Yali and Yang, Long and Knoll, Alois and Yang, Yaodong},
journal={Artificial Intelligence},
pages={103905},
year={2023},
publisher={Elsevier}
}