sekai

Environment catalog

Eleven environments are registered when sekai is imported. Build any of them with sekai.make(id); keyword arguments go to the environment’s constructor.

ID	Observation space	Action space	Horizon	Reward threshold
CartPole-v1	Box(4,)	Discrete(2)	500 (truncation)	475
MountainCar-v0	Box(2,)	Discrete(3)	200 (truncation)	−110
MountainCarContinuous-v0	Box(2,)	Box(-1, 1, (1,))	999 (truncation)	90
Pendulum-v1	Box(3,)	Box(-2, 2, (1,))	200 (truncation)	—
Acrobot-v1	Box(6,)	Discrete(3)	500 (truncation)	−100
Trading-v0	Box(lookback + 1,)	Discrete(len(positions))	episode_length = 252 (truncation)	—
Portfolio-v0	Dict(returns: Box(lookback, n), weights: Box(n + 1,))	Box(0, 1, (n + 1,)), or Box(-1, 1, (n,)) with shorts	episode_length = 252 (truncation)	—
MarketMaking-v0	Box(2,)	Box(0, max_offset, (2,))	n_steps = 200 (termination)	—
OptimalExecution-v0	Box(3,)	Box(0, 1, (1,))	n_steps = 25 (termination)	—
Market-v0	Dict(assets: MaskedBox(lookback, F) × max_assets, portfolio: Box(max_assets + 1,), account: Box(2,))	Box(0, 1, (max_assets + 1,)), or Box(-1, 1, (max_assets,)) with shorts	episode_length = 252 (truncation)	—
LimitOrderBook-v0	Dict(book: Box(levels, 4), trades: MaskedBox(3,) × max_trades, agent: Box(6,))	MultiDiscrete([K + 1, K + 1, 3])	n_steps = 300 (termination)	—

All observation and action Boxes are float32. For the classic-control environments the horizon is a TimeLimit added by make() (override it with make(id, max_episode_steps=N)); the finance environments end their own episodes, so make() adds no wrapper.

Trading-v0, Portfolio-v0, MarketMaking-v0 and OptimalExecution-v0 also have native vector implementations that step every sub-env with array operations, typically 20× faster than a SyncVecEnv. sekai.make_vec(id, n) uses them by default, and sekai list marks them [native vec].

import sekai

for spec in sekai.all_specs():
    env = sekai.make(spec.id)
    print(f"{spec.id:25s} {env.observation_space}")

CartPole-v1               Box([-4.8, -inf, -0.41887903, -inf], [4.8, inf, 0.41887903, inf], (4,), float32)
MountainCar-v0            Box([-1.2, -0.07], [0.6, 0.07], (2,), float32)
MountainCarContinuous-v0  Box([-1.2, -0.07], [0.6, 0.07], (2,), float32)
Pendulum-v1               Box([-1.0, -1.0, -8.0], [1.0, 1.0, 8.0], (3,), float32)
Acrobot-v1                Box([-1.0, -1.0, -1.0, -1.0, -12.566371, -28.274334], [1.0, 1.0, 1.0, 1.0, 12.566371, 28.274334], (6,), float32)
Trading-v0                Box(-inf, [inf, inf, inf, ..., inf, inf, 1.0], (31,), float32)
Portfolio-v0              Dict({returns: Box(-inf, inf, (30, 5), float32), weights: Box(0.0, 1.0, (6,), float32)})
MarketMaking-v0           Box([-1.0, 0.0], 1.0, (2,), float32)
OptimalExecution-v0       Box([0.0, 0.0, -inf], [1.0, 1.0, inf], (3,), float32)
Market-v0                 Dict({assets: MaskedBox(-inf, inf, (20, 2), max_size=10, float32), portfolio: Box(-inf, inf, (11,), float32), account: Box(-inf, inf, (2,), float32)})
LimitOrderBook-v0         Dict({book: Box(-inf, inf, (5, 4), float32), trades: MaskedBox(-inf, inf, (3,), max_size=16, float32), agent: Box(-inf, inf, (6,), float32)})

(A Box prints its overall minimum and maximum bound; see each environment’s page for the per-dimension ranges.)

Classic control

{ loading=lazy }

CartPole-v1

{ loading=lazy }

MountainCar-v0

{ loading=lazy }

MountainCarContinuous-v0

{ loading=lazy }

Pendulum-v1

{ loading=lazy }

Acrobot-v1

The standard control benchmarks, bit-identical to gymnasium, with NumPy-only RGB renderers.

Finance

Environment	You control	The baseline to beat
Trading-v0	a target position in one asset	buy and hold
Portfolio-v0	weights across assets and cash	equal weight, inverse volatility
MarketMaking-v0	bid and ask distances from the mid	Avellaneda–Stoikov closed-form quotes
OptimalExecution-v0	how much of a large order to sell each period	the Almgren–Chriss optimal schedule
Market-v0	weights across a universe whose members list and delist, with costs, leverage and financing	equal weight; an oracle that sees the market regime
LimitOrderBook-v0	bid and ask levels and market orders in a limit order book	inventory-skewed quotes

Trading-v0, Portfolio-v0 and Market-v0 run on simulated markets or on your own price history. LimitOrderBookMarket is a multi-agent order book (not registered; construct it from sekai.envs.finance). The components they are built from are documented in the finance toolkit.

Adding your own

Write a sekai.Env (guide), validate it with check_env and register it; it then shows up in sekai list and works with make(), make_vec() and the CLI like the built-ins.

This site is open source. Improve this page.