DM Control

The DeepMind Control Suite (Tassa et al., 2018) is a set of continuous control tasks, from a pendulum swing-up to a walking humanoid, and stable-envs has 17 of them as classes. dm_control builds, resets and restores each task. On CPU, MuJoCo steps the suite's own models in float64, and on CUDA, MuJoCo Warp steps them in float32. stable-envs ray casters draw the frames, except for the body-tracking tasks on CUDA, which MJWarp draws.

# needs the dmcontrol extra
import stable_envs as snv

with snv.DMCReacher(num_envs=8, device="cpu") as env:
    state = env.reset(seed=0)
    out = env.step(env.action_spec.sample(8))
    print(out["reward"].shape)     # torch.Size([8])
    print(env.render().shape)      # torch.Size([8, 84, 84, 3])

Install

The classes need the dmcontrol extra, which also installs Numba for the CPU renderer. On CUDA, the default physics and the body-tracking renderer need MuJoCo Warp, and Reacher's MJX physics needs MJX. The dmcontrol-gpu extra installs both; see what each family needs.

These classes default to device="cuda"; pass device="cpu" on a machine without a GPU.

Task

Each action dimension drives one actuator of the model, in the order of the MuJoCo model. Actions lie in [-1, 1] and are rescaled to each actuator's control range. Each step applies the action for two control steps of the task and returns the sum of their rewards, each in [0, 1]. Only Ball in cup defines a success criterion: success is 1 when the ball is in the cup.

Simulation

dm_control resets and restores each task. On CPU, every environment of a batch steps in one mujoco.batch call on a pool of threads. dmc_threads caps the number of threads, and a task with little MuJoCo work uses fewer: each thread gets about 0.1 ms of work per call, by a number of environments per thread that each task ships (see DM Control threads), and at least two environments. The number of threads doesn't change the state or the outputs. Build the environment in the process that steps it: the mujoco.batch threads don't survive fork().

State

The state of a task has three leaves, one for each dtype: dmc_<task>/float64, dmc_<task>/int64 and dmc_<task>/bool. For example, Reacher uses dmc_reacher/float64. Each field of the state is a slice of one of these leaves:

Leaf Fields
float64 qpos, qvel, time, warmstart, task/cumulative_reward, rng/cached_gaussian, activation (Quadruped), the target that the episode draws (task/target_pos or model/...) and the physics variation (physics/...)
int64 lifecycle/step_count, rng/keys, rng/position
bool lifecycle/reset_next_step, rng/has_gauss

Read a field with snv.get_field(state, "qpos"), and write a field with snv.set_field(state, "qpos", value). env.describe()["state_fields"] gives the shape and dtype of each leaf.

Physics backends

simulation_backend picks the physics when you construct the environment:

Value CPU device CUDA device
"auto" (default) native MuJoCo native MuJoCo for small batches, MuJoCo Warp for large ones
"native" native MuJoCo native MuJoCo on CPU, with state and outputs on CUDA
"mjwarp" error MuJoCo Warp
"mjx" error JAX/MJX, DMCReacher only

"mjwarp" and "mjx" raise an error without CUDA or without the GPU dependencies. With "auto" on CUDA, a batch of large_backend_min_envs or more environments steps on MuJoCo Warp, and a smaller batch steps on native MuJoCo. 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). Native MuJoCo computes in float64 and MuJoCo Warp in float32, so the two backends can give different trajectories. Set simulation_backend explicitly when you need the same physics on every machine.

# needs the dmcontrol-gpu extra; CUDA only
with snv.DMCReacher(num_envs=1, device="cuda") as env:
    print(env.simulation_backend)  # native: below the shipped switch point of 128

with snv.DMCReacher(num_envs=1, device="cuda", simulation_backend="mjwarp") as env:
    print(env.simulation_backend)  # mjwarp

On MuJoCo Warp, dm_control still resets the environments and owns their random streams, and the state keeps its shapes and dtypes, including the float64 physics columns. set_state followed by get_state returns the same state exactly. MuJoCo Warp does not repeat a contact solve bit for bit, so two rollouts from one state can differ after contacts. The shared WarpBatch backend (stable_envs/_execution/mjbatch.py) holds the Warp buffers and steps each control round in one captured CUDA graph. The adapter writes the model values of each environment into the per-environment model fields of MuJoCo Warp, and writes only the fields that changed. CheetahRun resets its environments on one dm_control owner per dmc_threads thread, and the other tasks on one owner.

Each control round follows dm_control's legacy_step: mj_step2, then n - 1 calls of mj_step, then mj_step1. RK4 tasks use n calls of mj_step, then mj_step1. After a round, the data hold the position and velocity stages of the new state and the acceleration-stage sensor values of the last integration. Contacts, constraints and the solver are those of the MuJoCo Warp 3.17 step, with no changes.

The step does not read values from the device. Nonfinite controls or physics raise dm_control.rl.control.PhysicsError one or two steps later. A contact or constraint buffer overflow raises BufferOverflowError in the same way. check_buffers() on the simulator raises the error immediately.

The table compares MuJoCo Warp with native CPU physics over 300 public calls (600 control rounds) from the same reset (seed=5), on 16 environments with slow sinusoidal actions. It gives the median over environments of the absolute return difference, the mean over the rollout of the largest qpos difference of the median environment, and the largest qpos difference.

Class Return, CPU (median) Return difference Mean qpos difference Largest qpos difference
DMCAcrobotBalance 0 0 6.5e-06 0.0056
DMCBallInCupCatch 0 0 7.8e-08 0.0011
DMCCartpoleBalance 177 0.00021 1.6e-05 0.0024
DMCCartpoleSwingup 0 0 7.2e-06 0.00055
DMCCheetahRun 3.84 1.7e-06 1.4e-07 1.4e-06
DMCFingerSpin 0 0 1.7e-06 6.2e-05
DMCFingerTurnEasy 0 0 1.3e-06 4.7e-05
DMCFingerTurnHard 0 0 1.3e-06 4.7e-05
DMCHopperHop 0 0 0.00052 1.9
DMCHumanoidWalk 0.66 4.9e-07 0.69 2.9
DMCManipulatorBringBall 2.8e-53 1.3e-58 0.0048 2
DMCPendulumSwingup 0 0 4.3e-07 3.3e-06
DMCPointMassEasy 3.8e-155 2.6e-158 8.3e-08 1.7e-06
DMCQuadrupedWalk 12.9 3.9e-05 3.4e-06 0.088
DMCReacher 4 0 7.7e-07 3.9e-06
DMCSwimmerSwimmer6 11.3 6e-06 3.5e-06 2e-05
DMCWalkerWalk 16.0 0.024 0.0098 2

In tasks with contacts, the float32 trajectories move apart and the difference grows over the rollout, but the returns stay close. Humanoid falls under these actions on both backends. Many tasks give a zero return under these actions, so their return difference shows little; the qpos columns cover them.

simulation_backend="mjx" keeps Reacher's JAX/MJX physics. On an L40S, MuJoCo Warp stepped Reacher faster than MJX at every batch from 1 to 4096, including resets. When the environment is built, stable-envs turns off JAX's GPU memory preallocation, so that Torch keeps its GPU memory. See JAX GPU memory.

Rendering

The renderers come in two groups:

  • The fixed-camera tasks (Acrobot, Ball in cup, Cartpole, Finger, Manipulator, Pendulum, Point mass and Reacher) use the analytic ray caster raycast3d. The tests compare their frames with MuJoCo's frames pixel by pixel.
  • The body-tracking tasks (Cheetah, Hopper, Humanoid, Quadruped, Swimmer and Walker) use cpu3d on CPU and MJWarp on CUDA. Their frames differ from MuJoCo's; see Limitations.

On CPU, one compiled Numba kernel draws the frames of each group, and the dmcontrol extra installs Numba. The kernel splits each frame into tiles of 8 × 8 pixels, and each tile tests its rays only against the moving geometry its bounds can reach, for shadow and mirror rays too. It then shades the pixels in the same order of float32 operations as the Torch renderer. backend="auto" (the default) picks Numba when it imports, and backend="torch" picks the Torch renderers. The two backends give the same frames except for rare pixels where Torch's CPU square root rounds differently from the IEEE square root the kernels use.

MuJoCo Warp renderer

backend="mjwarp" draws the frames on CUDA with the MuJoCo Warp ray tracer. Every class except the Manipulator classes accepts it. On CUDA, backend="auto" selects it for the body-tracking classes (Cheetah, Hopper, Humanoid, Quadruped, Swimmer and Walker). STABLE_ENVS_DMC_CUDA_RENDER=mjwarp also selects it for the other fixed-camera classes. On CPU, and for the Manipulator classes, backend="mjwarp" raises ValueError.

The renderer draws the geoms of the task. MuJoCo Warp does not draw sites and tendons, so the fixed-camera tasks add one geom for each site in groups 0-2 and one capsule for each segment of a drawn tendon. A kernel puts these geoms at the site poses and on the catenary of the tendon each frame. The catenary uses the gravity of each batch row. A pass replaces the sky with the source frame for cameras that keep one orientation. A second pass draws the floor reflection and the floor colors of each batch row.

Material colors, the light scale, the geom color leaves, the sphere radius leaves, gravity and the floor colors change per world. The field of view, visibility and shape leaves need a separate render model. The renderer groups the batch rows by these leaves and draws each group with its own context.

env.render(), env.render(state) and the frames of env.step run the same kinematics on the render rows of the State, so a frame is a function of the State.

render_options sets the speed and fidelity options. STABLE_ENVS_DMC_WARP_OPTIONS sets defaults as a JSON object.

Option Default Effect
shadows "all" "all": all lights that cast shadows do so. "one": only the first of them. "none": no shadow rays.
samples None Sub-pixel samples per axis. None uses supersample.
groups (0, 1, 2) The geom groups that the renderer draws.
bvh "refit" "refit" updates the bounds of the scene BVH each frame. "rebuild" builds the BVH again.
graph True Replays one captured CUDA graph per launch. False launches each kernel.
worlds None Worlds per launch. None uses the dmc_launch_rays and dmc_min_worlds tuning settings.
reflection True The floor reflection pass.
sky True The source sky for cameras that keep one orientation. False draws the MuJoCo Warp skybox.
segmentation, depth False Keep the geom id and the planar depth of each pixel.

graph, bvh and worlds change only the speed. The other options can change the pixels.

# needs the dmcontrol-gpu extra; CUDA only
env = snv.DMCCartpoleSwingup(num_envs=1024, device="cuda", backend="mjwarp", render_options={"shadows": "one"})

env.render_semantic(rows) returns the (objid, objtype) pair of each pixel in MuJoCo's mjtObj numbering and the planar depth along the optical axis, for compact render rows. Ids index the model of reference_env. The sky is -1.

Variation

Every class has material.*, floor.color and light.* variation leaves. Eleven tasks have camera.fovy_scale, and most have gravity.* and agent.* leaves, such as body densities or the cart mass. Cheetah, Hopper, Humanoid, Quadruped and Walker also vary floor.friction. env.physics_variation_leaves lists the leaves the simulator reads as physics.

Classes

Acrobot

DMCAcrobotBalance

Acrobot is a chain of two links with one actuator at the elbow. The task is to swing the free end up above the base and keep it there, with a sparse reward. The episode starts from random joint angles.

Random actions
Sampled variation (top left: default)
Property Value
Reference dm_control/acrobot-swingup_sparse
Physics native MuJoCo on CPU; MuJoCo Warp on CUDA
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (Numba)
Renderer on CUDA raycast3d (torch)
Default frame 84 × 84 × 3, uint8
Action float64 (1,), [-1, 1]
Clips rendered on CUDA at 160 px, then resized

State (1 top-level keys)

Key Shape dtype
dmc_acrobot 3 leaves, 640 values bool, float64, int64

Variation space (9 leaves)

Path Type Bounds or choices Default Sampled at reset
camera.fovy_scale Box (1,) [0.55, 1.8] 1
floor.color Box (2, 3) [0, 1] 6 values
gravity.x Box (1,) [-5, 5] 0
gravity.y Box (1,) [-5, 5] 0
gravity.z Box (1,) [-20, 0] -9.81
light.diffuse_scale Box (1,) [0.2, 1.8] 1
material.decoration Box (3,) [0, 1] [0.3, 0.5, 0.7]
material.grid Box (3,) [0, 1] 1
material.self Box (3,) [0, 1] [0.7, 0.5, 0.3]

Ball in cup

DMCBallInCupCatch

A cup on a planar arm holds a ball on a string. The agent swings the ball up and catches it in the cup.

Random actions
Sampled variation (top left: default)
Property Value
Reference dm_control/ball_in_cup-catch
Physics native MuJoCo on CPU; MuJoCo Warp on CUDA
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (Numba)
Renderer on CUDA raycast3d (torch)
Default frame 84 × 84 × 3, uint8
Action float64 (2,), [-1, 1]
Clips rendered on CUDA at 160 px, then resized

State (1 top-level keys)

Key Shape dtype
dmc_ball_in_cup 3 leaves, 649 values bool, float64, int64

Variation space (12 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.density Box (1,) [500, 1500] 1000
ball.density Box (1,) [500, 1500] 1000
ball.size Box (1,) [0.01, 0.05] 0.025
camera.fovy_scale Box (1,) [0.55, 1.8] 1
floor.color Box (2, 3) [0, 1] 6 values
gravity.x Box (1,) [-5, 5] 0
gravity.y Box (1,) [-5, 5] 0
gravity.z Box (1,) [-20, 0] -9.81
light.diffuse_scale Box (1,) [0.2, 1.8] 1
material.effector Box (3,) [0, 1] [0.7, 0.4, 0.2]
material.grid Box (3,) [0, 1] 1
material.self Box (3,) [0, 1] [0.7, 0.5, 0.3]

Cartpole

DMCCartpoleBalance

The pole starts near the upright pose. The agent moves the cart to keep it upright.

Random actions
Sampled variation (top left: default)
Property Value
Reference dm_control/cartpole-balance
Physics native MuJoCo on CPU; MuJoCo Warp on CUDA
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (Numba)
Renderer on CUDA raycast3d (torch)
Default frame 84 × 84 × 3, uint8
Action float64 (1,), [-1, 1]
Clips rendered on CUDA at 160 px, then resized

State (1 top-level keys)

Key Shape dtype
dmc_cartpole_balance 3 leaves, 641 values bool, float64, int64

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.cart_mass Box (1,) [0.5, 1.5] 1
camera.fovy_scale Box (1,) [0.55, 1.8] 1
floor.color Box (2, 3) [0, 1] 6 values
gravity.x Box (1,) [-5, 5] 0
gravity.y Box (1,) [-5, 5] 0
gravity.z Box (1,) [-20, 0] -9.81
light.diffuse_scale Box (1,) [0.2, 1.8] 1
material.decoration Box (3,) [0, 1] [0.3, 0.5, 0.7]
material.grid Box (3,) [0, 1] 1
material.self Box (3,) [0, 1] [0.7, 0.5, 0.3]

DMCCartpoleSwingup

The pole starts pointing down. The agent swings it up and keeps it upright.

Random actions
Sampled variation (top left: default)
Property Value
Reference dm_control/cartpole-swingup
Physics native MuJoCo on CPU; MuJoCo Warp on CUDA
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (Numba)
Renderer on CUDA raycast3d (torch)
Default frame 84 × 84 × 3, uint8
Action float64 (1,), [-1, 1]
Clips rendered on CUDA at 160 px, then resized

State (1 top-level keys)

Key Shape dtype
dmc_cartpole 3 leaves, 641 values bool, float64, int64

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.cart_mass Box (1,) [0.5, 1.5] 1
camera.fovy_scale Box (1,) [0.55, 1.8] 1
floor.color Box (2, 3) [0, 1] 6 values
gravity.x Box (1,) [-5, 5] 0
gravity.y Box (1,) [-5, 5] 0
gravity.z Box (1,) [-20, 0] -9.81
light.diffuse_scale Box (1,) [0.2, 1.8] 1
material.decoration Box (3,) [0, 1] [0.3, 0.5, 0.7]
material.grid Box (3,) [0, 1] 1
material.self Box (3,) [0, 1] [0.7, 0.5, 0.3]

Cheetah

DMCCheetahRun

A planar biped runs forward. The reward increases with the forward speed, up to 10 m/s.

Random actions
Sampled variation (top left: default)
Property Value
Reference dm_control/cheetah-run
Physics native MuJoCo on CPU; MuJoCo Warp on CUDA
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU cpu3d (Numba)
Renderer on CUDA mjwarp
Default frame 84 × 84 × 3, uint8
Action float64 (6,), [-1, 1]
Clips rendered on CUDA at 160 px, then resized

State (1 top-level keys)

Key Shape dtype
dmc_cheetah 3 leaves, 664 values bool, float64, int64

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.back_foot_density Box (1,) [500, 1500] 1000
agent.torso_density Box (1,) [500, 1500] 1000
floor.color Box (2, 3) [0, 1] 6 values
floor.friction Box (1,) [0, 1] 1
gravity.x Box (1,) [-5, 5] 0
gravity.y Box (1,) [-5, 5] 0
gravity.z Box (1,) [-20, 0] -9.81
light.diffuse_scale Box (1,) [0.2, 1.8] 1
material.grid Box (3,) [0, 1] 1
material.self Box (3,) [0, 1] [0.7, 0.5, 0.3]

Finger

DMCFingerSpin

A planar finger with two joints spins a hinged body as fast as possible.

Random actions
Sampled variation (top left: default)
Property Value
Reference dm_control/finger-spin
Physics native MuJoCo on CPU; MuJoCo Warp on CUDA
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (Numba)
Renderer on CUDA raycast3d (torch)
Default frame 84 × 84 × 3, uint8
Action float64 (2,), [-1, 1]
Clips rendered on CUDA at 160 px, then resized

State (1 top-level keys)

Key Shape dtype
dmc_finger_spin 3 leaves, 640 values bool, float64, int64

Variation space (7 leaves)

Path Type Bounds or choices Default Sampled at reset
camera.fovy_scale Box (1,) [0.55, 1.8] 1
floor.color Box (2, 3) [0, 1] 6 values
light.diffuse_scale Box (1,) [0.2, 1.8] 1
material.decoration Box (3,) [0, 1] [0.3, 0.5, 0.7]
material.effector Box (3,) [0, 1] [0.7, 0.4, 0.2]
material.grid Box (3,) [0, 1] 1
material.self Box (3,) [0, 1] [0.7, 0.5, 0.3]

DMCFingerTurnEasy

The finger turns the body until its tip points at a target. The target is large.

Random actions
Sampled variation (top left: default)
Property Value
Reference dm_control/finger-turn_easy
Physics native MuJoCo on CPU; MuJoCo Warp on CUDA
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (Numba)
Renderer on CUDA raycast3d (torch)
Default frame 84 × 84 × 3, uint8
Action float64 (2,), [-1, 1]
Clips rendered on CUDA at 160 px, then resized

State (1 top-level keys)

Key Shape dtype
dmc_finger_turn_easy 3 leaves, 642 values bool, float64, int64

Variation space (9 leaves)

Path Type Bounds or choices Default Sampled at reset
camera.fovy_scale Box (1,) [0.55, 1.8] 1
floor.color Box (2, 3) [0, 1] 6 values
light.diffuse_scale Box (1,) [0.2, 1.8] 1
material.decoration Box (3,) [0, 1] [0.3, 0.5, 0.7]
material.effector Box (3,) [0, 1] [0.7, 0.4, 0.2]
material.grid Box (3,) [0, 1] 1
material.self Box (3,) [0, 1] [0.7, 0.5, 0.3]
material.target Box (3,) [0, 1] [0.6, 0.3, 0.3]
visibility.target Discrete 2 choices from 0 1

DMCFingerTurnHard

This task is the same as DMCFingerTurnEasy with a smaller target.

Random actions
Sampled variation (top left: default)
Property Value
Reference dm_control/finger-turn_hard
Physics native MuJoCo on CPU; MuJoCo Warp on CUDA
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (Numba)
Renderer on CUDA raycast3d (torch)
Default frame 84 × 84 × 3, uint8
Action float64 (2,), [-1, 1]
Clips rendered on CUDA at 160 px, then resized

State (1 top-level keys)

Key Shape dtype
dmc_finger 3 leaves, 646 values bool, float64, int64
Variation space (15 leaves)
Path Type Bounds or choices Default Sampled at reset
agent.fingertip_density Box (1,) [500, 1500] 1000
agent.proximal_density Box (1,) [500, 1500] 1000
camera.fovy_scale Box (1,) [0.55, 1.8] 1
floor.color Box (2, 3) [0, 1] 6 values
light.diffuse_scale Box (1,) [0.2, 1.8] 1
material.decoration Box (3,) [0, 1] [0.3, 0.5, 0.7]
material.effector Box (3,) [0, 1] [0.7, 0.4, 0.2]
material.grid Box (3,) [0, 1] 1
material.self Box (3,) [0, 1] [0.7, 0.5, 0.3]
material.target Box (3,) [0, 1] [0.6, 0.3, 0.3]
spinner.color Box (3,) [0, 1] [0.7, 0.5, 0.3]
spinner.density Box (1,) [500, 1500] 1000
spinner.friction Box (1,) [0, 1] 0.1
target.shape Discrete 2 choices from 0 1
visibility.target Discrete 2 choices from 0 1

Hopper

DMCHopperHop

A planar one-legged body hops forward. The reward needs an upright torso and forward speed.

Random actions
Sampled variation (top left: default)
Property Value
Reference dm_control/hopper-hop
Physics native MuJoCo on CPU; MuJoCo Warp on CUDA
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU cpu3d (Numba)
Renderer on CUDA mjwarp
Default frame 84 × 84 × 3, uint8
Action float64 (4,), [-1, 1]
Clips rendered on CUDA at 160 px, then resized

State (1 top-level keys)

Key Shape dtype
dmc_hopper 3 leaves, 658 values bool, float64, int64

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.foot_density Box (1,) [500, 1500] 1000
agent.torso_density Box (1,) [500, 1500] 1000
floor.color Box (2, 3) [0, 1] 6 values
floor.friction Box (1,) [0, 1] 1
gravity.x Box (1,) [-5, 5] 0
gravity.y Box (1,) [-5, 5] 0
gravity.z Box (1,) [-20, 0] -9.81
light.diffuse_scale Box (1,) [0.2, 1.8] 1
material.grid Box (3,) [0, 1] 1
material.self Box (3,) [0, 1] [0.7, 0.5, 0.3]

Humanoid

DMCHumanoidWalk

A humanoid with 21 joints walks forward at 1 m/s.

Random actions
Sampled variation (top left: default)
Property Value
Reference dm_control/humanoid-walk
Physics native MuJoCo on CPU; MuJoCo Warp on CUDA
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU cpu3d (Numba)
Renderer on CUDA mjwarp
Default frame 84 × 84 × 3, uint8
Action float64 (21,), [-1, 1]
Clips rendered on CUDA at 160 px, then resized

State (1 top-level keys)

Key Shape dtype
dmc_humanoid 3 leaves, 719 values bool, float64, int64

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.right_lower_arm_density Box (1,) [500, 1500] 1000
agent.torso_density Box (1,) [500, 1500] 1000
floor.color Box (2, 3) [0, 1] 6 values
floor.friction Box (1,) [0, 1] 1
gravity.x Box (1,) [-5, 5] 0
gravity.y Box (1,) [-5, 5] 0
gravity.z Box (1,) [-20, 0] -9.81
light.diffuse_scale Box (1,) [0.2, 1.8] 1
material.grid Box (3,) [0, 1] 1
material.self Box (3,) [0, 1] [0.7, 0.5, 0.3]

Manipulator

DMCManipulatorBringBall

A planar arm with a gripper picks up a ball and brings it to a target position.

Random actions
Sampled variation (top left: default)
Property Value
Reference dm_control/manipulator-bring_ball
Physics native MuJoCo on CPU; MuJoCo Warp on CUDA
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (Numba)
Renderer on CUDA raycast3d (torch)
Default frame 224 × 224 × 3, uint8
Action float64 (5,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
dmc_manipulator 3 leaves, 670 values bool, float64, int64
Variation space (13 leaves)
Path Type Bounds or choices Default Sampled at reset
agent.hand_density Box (1,) [500, 1500] 1000
agent.upper_arm_density Box (1,) [500, 1500] 1000
camera.fovy_scale Box (1,) [0.55, 1.8] 1
floor.color Box (2, 3) [0, 1] 6 values
light.diffuse_scale Box (1,) [0.2, 1.8] 1
material.background Box (3,) [0, 1] 1
material.decoration Box (3,) [0, 1] [0.3, 0.5, 0.7]
material.effector Box (3,) [0, 1] [0.7, 0.4, 0.2]
material.grid Box (3,) [0, 1] 1
material.self Box (3,) [0, 1] [0.7, 0.5, 0.3]
material.target Box (3,) [0, 1] [0.6, 0.3, 0.3]
target_ball.shape Discrete 3 choices from 0 2
visibility.target_ball Discrete 2 choices from 0 1

Pendulum

DMCPendulumSwingup

The pendulum starts at a random angle. The agent swings it up and holds it upright with a limited torque.

Random actions
Sampled variation (top left: default)
Property Value
Reference dm_control/pendulum-swingup
Physics native MuJoCo on CPU; MuJoCo Warp on CUDA
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (Numba)
Renderer on CUDA raycast3d (torch)
Default frame 84 × 84 × 3, uint8
Action float64 (1,), [-1, 1]
Clips rendered on CUDA at 160 px, then resized

State (1 top-level keys)

Key Shape dtype
dmc_pendulum 3 leaves, 637 values bool, float64, int64

Variation space (10 leaves)

Path Type Bounds or choices Default Sampled at reset
camera.fovy_scale Box (1,) [0.55, 1.8] 1
floor.color Box (2, 3) [0, 1] 6 values
gravity.x Box (1,) [-5, 5] 0
gravity.y Box (1,) [-5, 5] 0
gravity.z Box (1,) [-20, 0] -9.81
light.diffuse_scale Box (1,) [0.2, 1.8] 1
material.decoration Box (3,) [0, 1] [0.3, 0.5, 0.7]
material.effector Box (3,) [0, 1] [0.7, 0.4, 0.2]
material.grid Box (3,) [0, 1] 1
material.self Box (3,) [0, 1] [0.7, 0.5, 0.3]

Point mass

DMCPointMassEasy

A point mass moves on a plane. The agent drives it to a target at the center.

Random actions
Sampled variation (top left: default)
Property Value
Reference dm_control/point_mass-easy
Physics native MuJoCo on CPU; MuJoCo Warp on CUDA
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (Numba)
Renderer on CUDA raycast3d (torch)
Default frame 84 × 84 × 3, uint8
Action float64 (2,), [-1, 1]
Clips rendered on CUDA at 160 px, then resized

State (1 top-level keys)

Key Shape dtype
dmc_point_mass_easy 3 leaves, 637 values bool, float64, int64

Variation space (7 leaves)

Path Type Bounds or choices Default Sampled at reset
camera.fovy_scale Box (1,) [0.55, 1.8] 1
floor.color Box (2, 3) [0, 1] 6 values
light.diffuse_scale Box (1,) [0.2, 1.8] 1
material.decoration Box (3,) [0, 1] [0.3, 0.5, 0.7]
material.grid Box (3,) [0, 1] 1
material.self Box (3,) [0, 1] [0.7, 0.5, 0.3]
material.target Box (3,) [0, 1] [0.6, 0.3, 0.3]

Quadruped

DMCQuadrupedWalk

A body with four legs walks forward.

Random actions
Sampled variation (top left: default)
Property Value
Reference dm_control/quadruped-walk
Physics native MuJoCo on CPU; MuJoCo Warp on CUDA
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU cpu3d (Numba)
Renderer on CUDA mjwarp
Default frame 224 × 224 × 3, uint8
Action float64 (12,), [-1, 1]
Clips rendered on CUDA at 224 px, then resized

State (1 top-level keys)

Key Shape dtype
dmc_quadruped 3 leaves, 716 values bool, float64, int64

Variation space (11 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.foot_back_left_density Box (1,) [500, 1500] 1000
agent.torso_density Box (1,) [500, 1500] 1000
floor.color Box (2, 3) [0, 1] 6 values
floor.friction Box (1,) [0, 1] 1
gravity.x Box (1,) [-5, 5] 0
gravity.y Box (1,) [-5, 5] 0
gravity.z Box (1,) [-20, 0] -9.81
light.diffuse_scale Box (1,) [0.2, 1.8] 1
material.effector Box (3,) [0, 1] [0.7, 0.4, 0.2]
material.grid Box (3,) [0, 1] 1
material.self Box (3,) [0, 1] [0.6, 0.3, 0.3]

Reacher

DMCReacher

A planar arm with two links moves its fingertip to a random target. It also has an MJX backend; see Physics backends.

Random actions
Sampled variation (top left: default)
Property Value
Reference dm_control/reacher-hard
Physics native MuJoCo on CPU; MuJoCo Warp on CUDA
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU raycast3d (Numba)
Renderer on CUDA raycast3d (torch)
Default frame 84 × 84 × 3, uint8
Action float64 (2,), [-1, 1]
Clips rendered on CUDA at 160 px, then resized

State (1 top-level keys)

Key Shape dtype
dmc_reacher 3 leaves, 642 values bool, float64, int64

Variation space (12 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.arm_density Box (1,) [500, 1500] 1000
agent.finger_density Box (1,) [500, 1500] 1000
camera.fovy_scale Box (1,) [0.55, 1.8] 1
floor.color Box (2, 3) [0, 1] 6 values
light.diffuse_scale Box (1,) [0.2, 1.8] 1
material.decoration Box (3,) [0, 1] [0.3, 0.5, 0.7]
material.effector Box (3,) [0, 1] [0.7, 0.4, 0.2]
material.grid Box (3,) [0, 1] 1
material.self Box (3,) [0, 1] [0.7, 0.5, 0.3]
material.target Box (3,) [0, 1] [0.6, 0.3, 0.3]
target.shape Discrete 2 choices from 0 1
visibility.target Discrete 2 choices from 0 0

Swimmer

DMCSwimmerSwimmer6

A planar swimmer with six links swims to a target in a viscous fluid.

Random actions
Sampled variation (top left: default)
Property Value
Reference dm_control/swimmer-swimmer6
Physics native MuJoCo on CPU; MuJoCo Warp on CUDA
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU cpu3d (Numba)
Renderer on CUDA mjwarp
Default frame 84 × 84 × 3, uint8
Action float64 (5,), [-1, 1]
Clips rendered on CUDA at 160 px, then resized

State (1 top-level keys)

Key Shape dtype
dmc_swimmer_swimmer6 3 leaves, 659 values bool, float64, int64

Variation space (7 leaves)

Path Type Bounds or choices Default Sampled at reset
floor.color Box (2, 3) [0, 1] 6 values
light.diffuse_scale Box (1,) [0.2, 1.8] 1
material.effector Box (3,) [0, 1] [0.7, 0.4, 0.2]
material.eye Box (3,) [0, 1] [0, 0.2, 1]
material.grid Box (3,) [0, 1] 1
material.self Box (3,) [0, 1] [0.7, 0.5, 0.3]
material.target Box (3,) [0, 1] [0.6, 0.3, 0.3]

Walker

DMCWalkerWalk

A planar biped walks forward at 1 m/s and keeps its torso upright.

Random actions
Sampled variation (top left: default)
Property Value
Reference dm_control/walker-walk
Physics native MuJoCo on CPU; MuJoCo Warp on CUDA
Devices CPU, CUDA
Requires CUDA no
Renderer on CPU cpu3d (Numba)
Renderer on CUDA mjwarp
Default frame 84 × 84 × 3, uint8
Action float64 (6,), [-1, 1]
Clips rendered on CUDA at 160 px, then resized

State (1 top-level keys)

Key Shape dtype
dmc_walker 3 leaves, 665 values bool, float64, int64

Variation space (11 leaves)

Path Type Bounds or choices Default Sampled at reset
agent.left_foot_density Box (1,) [500, 1500] 1000
agent.torso_density Box (1,) [500, 1500] 1000
floor.color Box (2, 3) [0, 1] 6 values
floor.friction Box (1,) [0, 1] 1
floor.rotation_y Box (1,) [-10, 10] 0
gravity.x Box (1,) [-5, 5] 0
gravity.y Box (1,) [-5, 5] 0
gravity.z Box (1,) [-20, 0] -9.81
light.diffuse_scale Box (1,) [0.2, 1.8] 1
material.grid Box (3,) [0, 1] 1
material.self Box (3,) [0, 1] [0.7, 0.5, 0.3]

Limitations

  • The body-tracking frames differ from MuJoCo's by a measured amount.
  • The body-tracking classes draw floor.color from a per-texel palette texture, and DMCWalkerWalk floor.rotation_y tilts the floor per batch row.
  • Native physics uses float64, and MuJoCo Warp and MJX use float32. They can produce different trajectories and rewards, and the gap grows on contact-heavy tasks. Use the same backend for every run you compare.
  • MuJoCo Warp uses the contacts, constraints and solver of MuJoCo Warp 3.17 with no changes. Physics backends gives the measured difference from native physics.
  • The tests compare against the upstream pins in pin-dmc: mujoco==3.17.0 and dm-control==1.0.50. The VISUAL_FIDELITY declarations of the body-tracking classes record the versions they were measured with.