You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
We provide a detailed description of the environment here.
Humanoid Strike
Humanoid Strike is a 3D environment with two simulated humanoid physics characters. Each character is equipped with a sword and shield with 37 degrees-of-freedom.
The game will be restarted if one agent goes outside the arena or the game reaches the maximum episode steps. We measure how much the player damaged the opponent and how much the player was damaged by the opponent in the terminated step to determine the winner.
Low-Level Observation Space
Index
Description
0
Height of the root from the ground.
1 - 48
Position of the body in the character’s local coordinate frame.
49 - 150
Rotation of the body in the character’s local coordinate frame.
151 - 201
Linear velocity of the root in the character’s local coordinate frame.
202 - 252
angular velocity of the root in the character’s local coordinate frame.
High-Level Observation Space
Index
Description
0 - 1
relative distance from the borderline
2 - 4
relative distance from the opponent
5 - 10
Rotation of the opponent's root in the character’s local coordinate frame.
11 - 13
Linear velocity of the opponent'root in the character’s local coordinate frame.
14 - 16
angular velocity of the opponent'root in the character’s local coordinate frame.
17 - 19
relative distance between ego agent and opponent's sword
20 - 22
Linear velocity of the opponent' sword in the character’s local coordinate frame.
23 - 25
relative distance between ego agent' shield and opponent's sword
26 - 28
relative velocity between ego agent' shield and opponent's sword
29 - 31
relative distance between ego agent' sword and opponent's torse
32 - 34
relative velocity between ego agent' sword and opponent's torse
35 - 37
relative distance between ego agent' sword and opponent's head
38 - 40
relative velocity between ego agent' sword and opponent's head
41 - 43
relative distance between ego agent' sword and opponent's right arm
44 - 46
relative distance between ego agent' sword and opponent's right thigh
47 - 49
relative distance between ego agent' sword and opponent's left thigh
Ant Sumo is a 3D environment with simulated physics that allows pairs of ant agents to compete against each other.
To win, the agent has to push the opponent out of the ring. Every agent has 100 hp . Each step, If the agent's body
touches the ground, its hp will be reduced by 1.The agent whose hp becomes 0 will be eliminated.
Observation Space
Index
Description
0 - 2
self pose
3 - 6
self rotation
7 - 9
self linear velocity
10 - 12
self angle velocity
13 - 20
self dof pos
21 - 28
self dof velocity
29 - 31
opponent pose
32 - 35
opponent rotation
36 - 37
self-opponent pose vector(x,y)
38
is self body touch ground
39
is opponent body touch ground
Action Space
Index
Description
0 - 7
self dof position
Rewards
The reward consists of two parts:sparse reward and dense reward.
Ant Battle is an expanded environment of Ant Sumo. It supports more than two agents competing against with
each other. The battle ring radius will shrink, the agent going out of the ring will be eliminated.
Observation Space
Index
Description
0 - 2
self pose
3 - 6
self rotation
7 - 9
self linear velocity
10 - 12
self angle velocity
13 - 20
self dof pos
21 - 28
self dof velocity
29
border radius-self dis to centre
30
border radius
31
is self body touch ground
32 - 34
opponent_1 pose
35 - 38
opponent_1 rotation
39 - 40
self-opponent_1 pose vector(x,y)
41 - 48
opponent_1 dof pose
49 - 56
opponent_1 dof velocity
57
border radius-opponent_1 dis to centre
58
is opponent_1 body touch ground
...
...
Action Space
Index
Description
0 - 7
self dof position
Rewards
The reward consists of two parts:sparse reward and dense reward.