Four-tank

Action space	Box((2,)), all actions in [-1, 1]
Observation space	Box((6,))
Tracked variable(s)	h1 (m), h2 (m)
Episode length	500 steps (500 s at 1 s per step)
Import	from target_gym import FourTank, FourTankParams
Cite as	four_tank-v1

Action space

Actions are normalised to [-1, 1] and mapped onto the plant's real actuator range inside the environment.

#	meaning	min	max
0	pump 1 voltage	-1	1
1	pump 2 voltage	-1	1

Observation space

6 values. Indices (0, 1) carry the tracked variable(s) that the reward scores.

Rewards

Mean of the two level-tracking scores.

Every environment in this suite scores on one contract: the reward is (tracking terms, multiplied) x (1 - weighted costs), bounded in [0, 1], and reaches 1 only while the target is held exactly. See Reward shaping.

Starting state

reset samples the initial condition and the target; state has 9 fields.

Episode end

Termination. See check_is_terminal.

Truncation. After 500 steps.

Baselines

A tuned PID ships with this environment.

Arguments

parameter	default
delta_t	1
max_steps_in_episode	500
g	9.81
gamma1	0.2
gamma2	0.2
k1	0.00085
k2	0.00095
a1	0.0035
a2	0.003
a3	0.002
a4	0.0025
A1	1
A2	1
A3	1
…	7 more, see the params dataclass