OGBench

OGBench (Park et al., 2024) is a benchmark for offline goal-conditioned reinforcement learning. stable-envs has 27 of its tasks as classes: 25 locomotion mazes, where a point mass, an ant or a humanoid moves through a maze to a goal cell, and 2 manipulation tasks, OGBCube and OGBScene, where a UR5e arm with a gripper moves objects on a table. MuJoCo steps every environment of a batch in one call on CPU and MuJoCo Warp steps them on CUDA. stable-envs draws the maze frames with its own ray caster, and Cube and Scene with MuJoCo OpenGL or MJWarp.

# needs the ogbench extra
import stable_envs as snv

with snv.PointMazeLargeNavigate(num_envs=8, device="cpu") as env:
    state = env.reset(seed=0)
    print(env.render().shape)      # torch.Size([8, 224, 224, 3])

Install

The mazes need the ogbench extra, and their default physics on CUDA also needs the mjwarp extra. Cube and Scene need mujoco==3.17.0 and the ogbench-manipulation extra, and on CUDA their default physics and supersample 2 or more also need ogbench-manipulation-gpu; see What each family needs.

The maze classes default to device="cuda"; pass device="cpu" on a machine without a GPU. Cube and Scene default to device="cpu".

Task

OGBench sets success when the agent reaches the goal, and the step output reports that value as success. In a maze, the body must be close to the goal position. In Cube and Scene, each object must match its goal state.

Each maze class name has three parts:

  • The body: PointMaze, AntMaze or HumanoidMaze.
  • The maze size: Medium, Large, Giant or Teleport. The teleport maze has cells that move the agent to another cell.
  • The dataset: Navigate, Stitch or Explore. The suffix selects the OGBench task and leaves the frames unchanged: classes that differ only in the suffix have the same clips. The exception is PointMazeLargeNavigate, which draws OGBench's state-observation scene, with a checker floor, white walls and the goal disc.

Simulation

The maze state holds the joint positions and velocities of the body, the clock and the solver warm start, the task and goal, and the generators. It has one tensor for each group: maze.physics.float64 changes on every step, and maze.host.float64, maze.host.int64, maze.host.uint8 and maze.host.bool change only at reset, restore and teleport. The fields are root.position, root.orientation and joints.position (the point body has only root.position), root.velocity, joints.velocity, simulator.time, simulator.warmstart, task.goal_pos, task.id and rng.*. Read a field by name with snv.get_field(state, "root.position").

Maze physics

One call steps all the environments of a batch, and simulation_backend selects the physics:

Value Device Physics Agreement with OGBench
"auto" (default) CPU mujoco.batch, float64 equal, bit for bit: state, observations, task outputs, resets and RNG streams
"auto" (default) CUDA mujoco.batch for small batches, MuJoCo Warp for large ones as the selected physics
"native" CPU or CUDA mujoco.batch, float64 equal, bit for bit
"mjwarp" CUDA MuJoCo Warp, float32 equal resets, goals, tasks, teleports and RNG streams; approximate physics

With "auto" on CUDA, a batch of large_backend_min_envs or more environments steps on MuJoCo Warp. When the setting is not set, the switch point comes from the tuning cache that stable-envs tune writes, else from the value the class ships, so without tuning the same batch size selects the same backend on every machine (Physics backends). The outputs are on the device of the environment. On CUDA, mujoco.batch moves the outputs and the state of a step to the device in one copy.

On CPU, mujoco.batch steps the physics on its own threads, and the step does its Torch operations on one thread. Idle Torch threads would otherwise keep the CPU cores busy while the mujoco.batch threads wait for them. The frames and the state do not change with the thread count.

# CUDA only; needs the ogbench and mjwarp extras
import stable_envs as snv

with snv.AntMazeMediumNavigate(num_envs=64, device="cuda") as env:
    print(env.simulation_backend)  # native: 64 is below the AntMaze switch point, 128

MuJoCo Warp agrees with MuJoCo to float32 rounding on a step where no contact starts or stops. When contacts start or stop, the contact forces of the two solvers differ and the trajectories move apart. After 200 steps of smooth actions from the same state on 64 environments, the median distance between the xy positions of MuJoCo Warp and OGBench was:

Body Median xy distance after 200 steps
PointMaze less than 1e-5 (largest: 0.13 after wall contacts)
AntMaze 0.05
HumanoidMaze 0.23 (0.11 when only the start state is rounded to float32 in MuJoCo)

A PointMaze agent that follows the OGBench subgoal oracle reaches the goal at the same step on both backends.

On MuJoCo Warp, PointMaze integrates with Euler, where OGBench uses RK4. The point's step moves the position by the action and sets the velocity to zero, so away from walls the two integrators give the same positions, and only the response to a wall contact differs: over 300 steps of smooth actions on 32 environments, the largest xy distance from OGBench was 0.13 (2e-5 with RK4). RK4 runs four forward passes in each substep, so the Euler step is about four times faster.

Cube and Scene physics

simulation_backend picks the physics:

Value CPU device CUDA device
"auto" (default) mujoco.batch mujoco.batch for small batches, MuJoCo Warp for large ones
"native" mujoco.batch mujoco.batch
"mjwarp" error MuJoCo Warp

With "auto" on CUDA, a batch of large_backend_min_envs or more environments steps on MuJoCo Warp. When the setting is not set, the switch point comes from the tuning cache that stable-envs tune writes, else from the value the class ships, so without tuning the same batch size selects the same backend on every machine (Physics backends).

With two or more environments, mujoco.batch steps every environment in one native call on up to 8 threads. One environment steps on one MjData with OGBench's own mj_step calls. The step does the operations of OGBench's step in the same order: the end-effector controller and its inverse kinematics, the 25 mj_step substeps, mj_rnePostConstraint and the task's post_step. The state and the outputs equal OGBench's bit for bit, and a CUDA environment on mujoco.batch returns its state on CUDA. The default model has the OGBench physics, with a cube size of 0.02.

MuJoCo Warp steps the substeps of every environment in one captured CUDA graph, in float32, and one Warp kernel computes the controller and its inverse kinematics in float64. The state has the same fields as on CPU, and its physics fields hold the float32 values. On the scripted OGBench Cube oracle, each episode ends with the same success at the same first-success step as OGBench (see tests/gpu/test_ogbench_manipulation_warp.py). Over a few hundred steps, contacts can make the cube positions differ from OGBench.

MuJoCo Warp doesn't test the links of the Robotiq gripper against each other for contact. The pads keep their contacts with each other and with all other geoms. The linkage has one degree of freedom and the pads touch before two links can, so these pairs make no contacts in OGBench either, and the step is faster without their mesh-mesh tests.

The MuJoCo Warp step doesn't synchronize with the host. When a Scene cube leaves the workspace, OGBench puts it at a new position drawn from the environment's generator. The host draws the next two of these positions for each environment before the step, and the device uses them. The host draws again for an environment after it reads that the device used them, and then prints OGBench's message. If an environment needs a third position before that read, the device uses the second one again.

Cube and Scene state

The Cube and Scene state has one leaf for each section and dtype, such as physics/float64. Each leaf holds the fields of its section side by side, for example physics/qpos, physics/qvel and physics/mocap_pos. Read and write a field by its name with snv.get_field and snv.set_field:

import stable_envs as snv

with snv.make("OGBCube", num_envs=2) as env:
    state = env.reset(seed=0)
    qpos = snv.get_field(state, "physics/qpos")  # (2, 21) float64
    size = snv.get_field(state, "variation/current/cube/size")  # also "variation.current.cube.size"

A variation leaf name such as cube.size becomes the field path cube/size. env.state_spec and describe() list the leaves.

Cube and Scene resets

A reset is the OGBench reset: resets, restores and variations run OGBench's own code on one OGBench object per environment, and the physics backend then loads the result. On Linux, a batch of two or more environments resets in up to workers worker processes, which reset the environments at the same time. Set worker_min_envs to change the smallest batch that uses them. OGBench samples the task from NumPy's global RNG, and each environment in the batch keeps its own copy of that RNG in state (rng/task). A seeded reset seeds the copy the way np.random.seed(seed) would with that environment's seed, and an unseeded reset continues it. OGBench samples the two settling actions of a reset from an unseeded action space. stable-envs samples them from the seeded rng/action generator instead. Restoring a state therefore also fixes the next reset.

Rendering

The analytic ray caster raycast3d draws the maze, the body and the floor from the state on CPU or CUDA. On CPU, the render uses the Torch thread count (torch.set_num_threads). It draws batches of fewer than 256 frames on one thread when the scene has no tracking light, which is every class except PointMazeLargeNavigate.

On CPU, PointMazeLargeNavigate draws the tracking light on the static pixels with Numba kernels that do all the frames of a render in one pass. The square roots and the powers of the light stay Torch operations on the same tensors as the Torch path, thus the frames are equal to the frames of the Torch path, byte for byte. Without Numba, the Torch path draws the light. The ogbench extra installs Numba.

Cube and Scene render backends

The frames have the look of the OGBench pixel tasks visual-cube-single-v0 and visual-scene-v0: the front_pixels camera, a purple gripper, a see-through arm and no goal cube.

The variation leaf agent.opacity sets the opacity of the arm. At the default value 0, the arm has the alphas of the OGBench pixel tasks. At 1, the arm is opaque, as in the OGBench state tasks and in swm's swm/OGBCube-v0 and swm/OGBScene-v0. The variation leaf target.opacity shows the goal cube. At the default value 0, the goal cube is hidden, as in the OGBench pixel tasks. At 1, the goal cube has alpha 0.2, as with OGBench's visualize_info=True and in swm's environments. Set the leaves with set_variation, then call reset. Each value of agent.opacity and target.opacity compiles its own render model on the MJWarp backends.

Device and setting backend="auto" draws with Frames
CPU the MuJoCo rasterizer in Numba, as backend="numba" differ from MuJoCo OpenGL; a draw of the same state gives the same frame each time, in any batch
CUDA, supersample=1 NVIDIA OpenGL the driver can change a pixel by 1 or 2 levels (of 255) between two draws of the same state, and so can the OpenGL reference
CUDA, supersample 2 or more two-pass MJWarp, as backend="mjwarp_fidelity" can differ from MuJoCo in lighting and transparency
  • backend="numba" is the shared MuJoCo rasterizer. It follows MuJoCo's OpenGL renderer: the same meshes and primitive tessellations, per-vertex lighting, the 4096 shadow map of the scene light, the floor texture with its mipmaps, the skybox, the see-through arm and 4 samples per pixel. It draws the arm meshes with one vertex per grid cell of 2 mm. On CPU, the first render after an install compiles the Numba kernels in about 1 to 2 minutes, and Numba keeps the compiled code in a cache for later processes. On CUDA, backend="numba" raises ValueError. render_semantic(rows) gives the segmentation and the planar depth of compact render rows.
  • Without Numba, the CPU default is software OpenGL (backend="opengl"). The CUDA default at supersample=1 is NVIDIA OpenGL.
  • backend="opengl" draws the frames of the source. On CPU, these frames are bit-exact. On CUDA, the NVIDIA driver can change a pixel by 1 or 2 levels (of 255) between two draws of the same state, and so can the OpenGL reference.
  • CPU OpenGL rendering refuses a hardware GL context. For backend="opengl", configure Mesa software OpenGL before you import MuJoCo.
  • The default supersample is 1 on every device. An explicit supersample argument keeps its value, and so does the sampling setting stored in a render state you pass in.
  • backend="mjwarp" needs CUDA.
  • On Linux, a batch of states draws in OpenGL worker processes, at most ogb_gl_workers. On CPU, the workers start when there are at least 2 of them and at least as many states as workers. On CUDA, there is one worker for every 8 states, and the workers start from 16 states. On CPU, the frames are the same as from one process. On CUDA, each worker has its own OpenGL context, so the frames can differ from one process by the driver variation in the table above: a pixel can change by 1 level, and the mean squared difference was below 2e-3 (on the 0-255 scale) on RTX A5000 and A5500 hosts.
  • light.intensity, camera.angle_delta, cube.size and Scene lock_color are per-world inputs to one MJWarp context per family. Changing them adds no contexts or compiled models.

Variation

The maze variation leaves set the colors of the body, the walls and the floor, the camera, the light and the body's agent.scale. Cube and Scene vary colors, the opacity of the arm and of the goal cube, the camera, the light, the cube size and the start and goal poses of the cube.

In Cube and Scene, the arm starts at the position OGBench samples. It starts at the variation leaf agent.ee_start_position only when the reset options name this leaf, or when set_variation sets it to a non-default value.

Classes

PointMaze

PointMazeMediumNavigate

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/pointmaze-medium-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (2,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 672 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

PointMazeMediumStitch

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/pointmaze-medium-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (2,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 672 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

PointMazeLargeNavigate

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/pointmaze-large-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (2,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 672 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 1

PointMazeLargeStitch

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/pointmaze-large-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (2,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 672 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

PointMazeGiantNavigate

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/pointmaze-giant-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (2,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 672 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

PointMazeGiantStitch

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/pointmaze-giant-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (2,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 672 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

PointMazeTeleportNavigate

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/pointmaze-teleport-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (2,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 672 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

PointMazeTeleportStitch

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/pointmaze-teleport-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (2,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 672 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

AntMaze

AntMazeMediumNavigate

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/antmaze-medium-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (8,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 709 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

AntMazeMediumStitch

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/antmaze-medium-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (8,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 709 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

AntMazeMediumExplore

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/antmaze-medium-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (8,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 709 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

AntMazeLargeNavigate

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/antmaze-large-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (8,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 709 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

AntMazeLargeStitch

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/antmaze-large-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (8,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 709 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

AntMazeLargeExplore

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/antmaze-large-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (8,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 709 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

AntMazeGiantNavigate

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/antmaze-giant-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (8,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 709 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

AntMazeGiantStitch

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/antmaze-giant-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (8,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 709 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

AntMazeTeleportNavigate

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/antmaze-teleport-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (8,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 709 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

AntMazeTeleportStitch

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/antmaze-teleport-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (8,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 709 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

AntMazeTeleportExplore

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/antmaze-teleport-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (8,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 709 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

HumanoidMaze

HumanoidMazeMediumNavigate

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/humanoidmaze-medium-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (21,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 748 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

HumanoidMazeMediumStitch

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/humanoidmaze-medium-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (21,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 748 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

HumanoidMazeLargeNavigate

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/humanoidmaze-large-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (21,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 748 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

HumanoidMazeLargeStitch

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/humanoidmaze-large-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (21,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 748 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

HumanoidMazeGiantNavigate

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/humanoidmaze-giant-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (21,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 748 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

HumanoidMazeGiantStitch

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/humanoidmaze-giant-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (torch)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float32 (21,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
maze 5 leaves, 748 values bool, float64, int64, uint8

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
agent.scale Box (1,) [0.25, 6] 1
camera.distance_scale Box (1,) [0.6, 1.8] 1
camera.elevation Box (1,) [-90, -35] -90
floor.color Box (2, 3) [0, 1] 6 values
floor.pixel_encoding Discrete 2 choices from 0 0
floor.tint Box (3,) [0, 1] 1
light.intensity Box (1,) [0, 1] 1
rendering.decolor_floor Discrete 2 choices from 0 0
walls.color Box (3,) [0, 1] 0.6

Manipulation

OGBCube

The arm moves one cube to a goal pose. This is the OGBench cube-single task, and the constructor argument reward_task_id picks the evaluation task the reward uses.

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/visual-cube-single-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU MuJoCo rasterizer (Numba); opengl (MuJoCo, software) without Numba
Renderer on CUDA opengl (MuJoCo, NVIDIA) at supersample 1; mjwarp at 2 or more
Default frame 224 × 224 × 3, uint8
Action float32 (5,), [-1, 1]
Clips rendered on CPU at 224 px, then resized

State (9 top-level keys)

Key Shape dtype
compatibility 1 leaves, 32 values uint8
physics 2 leaves, 255 values bool, float64
controller 1 leaves, 12 values float64
output 1 leaves, 47 values float64
task 3 leaves, 37 values bool, float64, int64
lifecycle 1 leaves, 1 values bool
variation 1 leaves, 54 values float64
model 1 leaves, 30 values float64
rng 3 leaves, 1409 values float64, int64, uint8
Variation space (13 leaves)
Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.61, 0.28, 0.82]
agent.ee_start_position Box (3,) [0.25, -0.35, 0.2] to [0.6, 0.35, 0.35] [0.425, 0, 0.275]
agent.opacity Box (1,) [0, 1] 0
camera.angle_delta Box (1, 2) [-10, 10] 0
cube.color Box (1, 3) [0, 1] [0.96, 0.26, 0.33]
cube.goal_position Box (1, 2) [0.3, -0.3] to [0.55, 0.3] [0.3, 0]
cube.goal_yaw Box (1,) [0, 6.28319] 0
cube.size Box (1,) [0.01, 0.03] 0.02
cube.start_position Box (1, 2) [0.3, -0.3] to [0.55, 0.3] [0.3, 0]
cube.start_yaw Box (1,) [0, 6.28319] 0
floor.color Box (2, 3) [0, 1] 6 values
light.intensity Box (1,) [0, 1] 0.7
target.opacity Box (1,) [0, 1] 0

OGBScene

The scene has a cube, a drawer, a window and two button locks. The agent changes the objects to match a goal configuration. A locked object stays in place until the agent presses its button.

Random actions
Sampled variation (top left: default)
Property Value
Reference ogbench/visual-scene-v0
Physics source simulator on CPU
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU MuJoCo rasterizer (Numba); opengl (MuJoCo, software) without Numba
Renderer on CUDA opengl (MuJoCo, NVIDIA) at supersample 1; mjwarp at 2 or more
Default frame 200 × 200 × 3, uint8
Action float32 (5,), [-1, 1]
Clips rendered on CPU at 200 px, then resized

State (10 top-level keys)

Key Shape dtype
compatibility 1 leaves, 32 values uint8
physics 2 leaves, 343 values bool, float64
controller 1 leaves, 42 values float64
output 1 leaves, 55 values float64
task 3 leaves, 57 values bool, float64, int64
lifecycle 1 leaves, 1 values bool
scene 2 leaves, 8 values float64, int64
variation 1 leaves, 66 values float64
model 1 leaves, 30 values float64
rng 3 leaves, 1443 values float64, int64, uint8
Variation space (14 leaves)
Path Type Bounds or choices Default Sampled at reset
agent.color Box (3,) [0, 1] [0.61, 0.28, 0.82]
agent.ee_start_position Box (3,) [0.25, -0.2, 0.2] to [0.6, 0.2, 0.35] [0.425, 0, 0.275]
agent.opacity Box (1,) [0, 1] 0
camera.angle_delta Box (1, 2) [-5, 5] 0
cube.color Box (1, 3) [0, 1] [0.96, 0.26, 0.33]
cube.goal_position Box (1, 2) [0.3, -0.07] to [0.45, 0.18] [0.375, 0.055]
cube.goal_yaw Box (1,) [0, 6.28319] 0
cube.size Box (1,) [0.01, 0.03] 0.02
cube.start_position Box (1, 2) [0.3, -0.07] to [0.45, 0.18] [0.375, 0.055]
cube.start_yaw Box (1,) [0, 6.28319] 0
floor.color Box (2, 3) [0, 1] 6 values
light.intensity Box (1,) [0, 1] 0.7
lock_color Box (2, 3) [0, 1] 6 values
target.opacity Box (1,) [0, 1] 0

Limitations

  • Cube and Scene are compared with frames from MuJoCo's OpenGL renderer on the model stable-envs compiles, not with a default OGBench install. The fidelity record of the NVIDIA OpenGL path compares two stable-envs OpenGL instances with the same compiled appearance.
  • With supersample 2 or more, MJWarp draws the CUDA frames. It is measured against the same stable-envs OpenGL path, but visual_fidelity holds only the values of backend="opengl".
  • HumanoidMazeGiantNavigate on CUDA isn't reproducible across processes on at least one GPU.
  • MuJoCo Warp orders contacts with atomic operations, so two runs from one state can differ by rounding. Use simulation_backend="native" when you need results equal to OGBench.
  • The default CPU numba frames differ from the backend="opengl" frames that visual_fidelity records. Only backend="opengl" gives the frames of the source.