Embedded Files
MuJoCo
MuJoCo
Agents
Agents
MuJoCo ( Multi Joints with Contacts ) is a free, open-source physics engine tailored for research in robotics, biomechanics, graphics, and more, offering exceptional speed, accuracy, and modeling capabilities.
Within this engine, we consider two sets of maze environments from the Gymnasium framework : PointMaze and AntMaze. The Point agent is a point mass controlled by applying forces along X and Y axis allowing it to move freely across the plane. In contrast the Ant agent is a more complex articulated quadruped robot, controlled through torques actuation.
The datasets are obtained through the Minari API.
Ant agent.
Point agent.
Mazes
Mazes
AntMaze : U-Maze
Dataset of 1000 expert episodes.
AntMaze : Medium
Dataset of 1000 expert episodes.
AntMaze : Large
Dataset of 1000 expert episodes.
PointMaze : U-Maze
Dataset of 1000 expert episodes.
PointMaze : Medium
Dataset of 1000 expert episodes.
PointMaze : Large
Dataset of 1000 expert episodes.
Tasks
Tasks
Default ( 3 tasks per env. )
Default ( 3 tasks per env. )
{env}-{maze}-normal
Topological Changes ( 9 tasks per env. )
Topological Changes ( 9 tasks per env. )
{env}-{maze}-normal
{env}-{maze}-shift_map_1
{env}-{maze}-shift_map_2
Kinematic Changes ( 15 tasks per env. )
Kinematic Changes ( 15 tasks per env. )
{env}-{maze}-normal
{env}-{maze}-inverse_actions
{env}-{maze}-permute_actions
{env}-{maze}-rescale_actions_1
{env}-{maze}-rescale_actions_2
Godot
Godot
Agent
Agent
Godot agent : upon generating trajectories both the human user and the learning agent use 'Z' 'Q' 'S' 'D' 'SPACE' keys and the mouse to control the agent.
Raycast vision of the agent : alongside position and orientation information, the learning agent uses depth map images obtained using 11 × 11 raycasts at 180°.
We introduce a video-game like 3D navigation environments, implemented with Godot and offering diverse tasks with explainable spatial challenges.
There are two families of mazes : SimpleTown ones are relatively simple, with a size of 30 × 30 meters, the starting positions are randomly sampled on one side, and the goal positions on the other side ; AmazeVille ones, with a size of 60 × 60 meters, have a finite set of start and goal positions, and include two subsets of maps : some with high blocks, i.e. not jumpable obstacles ; others with low blocks, i.e. jumpable ones.
In the Godot-based environments, the episodes datasets were sampled manually over approximately 10 hours, resulting in 100 episodes for each AmazeVille maze and 250 episodes for each SimpleTown maze.
Mazes
Mazes
SimpleTown : Base
Dataset of 250 expert episodes.
SimpleTown : XOO
Dataset of 250 expert episodes.
SimpleTown : OXO
Dataset of 250 expert episodes.
SimpleTown : OOX
Dataset of 250 expert episodes.
SimpleTown : OOO
Dataset of 250 expert episodes.
SimpleTown : OXX
Dataset of 250 expert episodes.
SimpleTown : XOX
Dataset of 250 expert episodes.
SimpleTown : XXO
Dataset of 250 expert episodes.
AmazeVille : HOOO
Dataset of 100 expert episodes.
AmazeVille : HOOX
Dataset of 100 expert episodes.
AmazeVille : HXOO
Dataset of 100 expert episodes.
AmazeVille : HXOX
Dataset of 100 expert episodes.
AmazeVille : LOOO
Dataset of 100 expert episodes.
AmazeVille : LOOX
Dataset of 100 expert episodes.
AmazeVille : LXOO
Dataset of 100 expert episodes.
AmazeVille : LXOX
Dataset of 100 expert episodes.
( Brief ) Dataset Analysis
( Brief ) Dataset Analysis
Paths - SimpleTown
Paths - AmazeVille
Positions Densities - SimpleTown
Positions Densities - AmazeVille
Page updated
Google Sites
Report abuse