Environments

StableTested and in active use. Breaking changes will be called out. DeepMind Control, the robot ports and LIBERO are parity-gated against MuJoCo but newer to training — see their pages.

Thirteen families, from a 5×5 grid to dm_control’s dog. Everything except the Gymnasium wrappers is Mojo: the tabular grids, the classic-control dynamics, the 2D and 3D physics, the robot and LIBERO scenes, the arcade engines and the Atari 2600 emulator. That is what lets an environment step inside the same GPU pipeline that trains on its output.

HalfCheetahSAC · physics3d

HumanoidSAC · physics3d

LunarLanderPPO · physics2d

CarRacingRainbow · physics2d

Antphysics3d

Pongnative GPU engine

The families

Family	Environments	Engine	GPU batch
Tabular	GridWorld · FrozenLake · CliffWalking · Taxi	grid	—
Classic control	CartPole · MountainCar · Acrobot · Pendulum	Gymnasium-matching	partial
2D physics	LunarLander · BipedalWalker · CarRacing	physics2d	yes
3D locomotion	HalfCheetah · Hopper · Walker2d · Ant · Swimmer · Humanoid · pendula	physics3d	yes
3D manipulation	Reacher · Pusher · Sawyer	physics3d	yes
DeepMind Control	18 suite domains, 49 tasks · 13 manipulation tasks	physics3d	15 domains
Robots	SO-ARM100 · SO-ARM101 · Unitree G1 · Franka Panda	physics3d	yes (G1: NVIDIA only)
LIBERO	23 task families on the Panda	physics3d	yes
Arcade	Pong · Breakout · Space Invaders	native GPU	yes
Atari 2600	103 ROMs	6502 emulation	CPU only
Board games	Tic-Tac-Toe · Connect Four · Go · Chess	—	yes
Procgen & Craftax	16 Procgen games · Craftax Classic and full	native	yes
Gymnasium	20+ wrappers	Python interop	—

The SO-101 and LIBERO scenes are also where the task layer lives: families of scenes with goals, placement rules and resets, rather than one hand-written environment per task.

Full reference

Dimensions below are read from the source, not from the Gymnasium documentation. Where noeira differs from Gymnasium’s defaults, it is noted.

Tabular

Environment	States	Actions	Reward
GridWorld	25 (5×5)	4	−1/step, +10 goal
FrozenLake	16 (4×4)	4	+1 goal (sparse)
CliffWalking	48 (4×12)	4	−1/step, −100 cliff
Taxi	500	6	+20 dropoff, −10 illegal

Classic control

Environment	Obs	Actions	Notes
CartPole	4	2 discrete	GPU batch
MountainCar	2	3 discrete	tile-coding friendly
Acrobot	6	3 discrete	RK4 integration
Pendulum	3	1 continuous	V1 CPU · V2 GPU batch

2D physics — physics2d

Environment	Obs	Actions	Notes
LunarLander	8	4 discrete or 2 continuous	terrain + flame particles
BipedalWalker	24	4 continuous	revolute joints, 10-ray LIDAR
CarRacing	13 clean or pixels	3 continuous or 5 discrete	slip-based tire model

3D — physics3d

Environment	Obs	Actions	Frame skip
HalfCheetah	17	6	5
Hopper	11	3	4
Walker2d	17	6	4
Ant	27	8	5
Swimmer	8	2	4
Humanoid	45	17	5
HumanoidStandup	45	17	5
InvertedPendulum	4	1	2
InvertedDoublePendulum	9	1	5
Reacher	10	2	2
Pusher	23	7	5

Humanoid is 45-dimensional, not 376

Gymnasium’s Humanoid-v5 observation includes contact forces and inertia terms, which push it to 348–376 dimensions. noeira’s Humanoid observes (NQ − 2) + NV = 22 + 23 = 45 — position and velocity only. Older copies of the project README and slide deck quote 376; the source is authoritative and the GPU example reads HumanoidModel.OBS_DIM # 45.

DeepMind Control, robots — physics3d

Environment	Obs	Actions	Episode
dm_control suite (walker, cheetah, humanoid, dog, …)	per task	per task	1000 steps, reward in [0, 1] per step
dm_control manipulation (Jaco arm and hand, bricks)	per task	9 actuators	250 steps
SO-ARM100 / SO-ARM101 reach	27	6 joint position targets, in [−1, 1] per joint	500 steps
Unitree G1	64 proprioceptive + 463 privileged	29 PD targets	500 steps at 50 Hz

See DeepMind Control and Robots for the per-task tables.

Arcade — native GPU engines

Two observation modes each: a compact game-state vector (OBS_MODE=0) or a 4×84×84 grayscale frame stack (OBS_MODE=1).

Environment	Clean obs	Actions
Pong	6	3
Breakout	7	4
Space Invaders	8	4

Where the dimensions live

A recurring source of confusion, so worth stating plainly: for the physics3d environments the dimensions are on the model, not the config.

comptime OBS_DIM = HumanoidModel.OBS_DIM      # 45
comptime ACT_DIM = HumanoidModel.ACTION_DIM   # 17

HalfCheetah and Ant additionally expose them on their config (HalfCheetahConfig.OBS_DIM), which is what the older examples use. Both resolve to the same compile-time constant. If one does not exist, try the other.

Choosing one

Learning the framework — GridWorld or CartPole. No GPU, seconds to run, obvious correct behaviour.

Testing a continuous-control algorithm — Pendulum first (3 dimensions, fast), then HalfCheetah (no early termination, so returns are smooth and easy to compare).

Testing exploration — MountainCar or FrozenLake. Both have sparse or deceptive reward that punishes greedy agents.

Stressing the physics engine — Humanoid. 45 observations, 17 actuators, and contact-rich dynamics that expose solver problems immediately.

Comparing against published continuous-control numbers — the DeepMind Control ports, whose returns are bounded by 1000 and comparable across tasks. They are parity-gated against MuJoCo but none has a training result documented yet.

A robot you can own — SO-ARM101, which has a real-hardware path.

Pixel-based agents — the arcade engines in OBS_MODE=1. Native and GPU-batched, so they are far faster than the Atari emulator for training; the emulator exists to validate them.

Next

• Writing your own environment — the traits to implement
• Traits — the trait hierarchy in detail
• CPU and GPU paths — how batched environments work