Environment reference

Twenty-one environments, every one a target MDP: the objective is to reach and hold a subset of the state space against disturbances, not to reach a goal and stop.

Every environment implements the same interface, so the table below is the whole of what differs between them:

• Observation / Action are the array shapes.
• Tracked is the quantity the target applies to, exposed at env.obs_value_index with its setpoint at env.obs_target_index, and named by env.tracked_names.
• PID / MPC say whether a baseline ships. Where one does not, the registry records why in baselines_note rather than leaving it silent.
• Physics links the environment's contract: sourced parameters, published validation targets asserted by tests, and quantified known deviations.

Load any of them by registry name:

from target_gym.registry import REGISTRY

spec = REGISTRY["boiler_drum"]
env, params = spec.make_env(), spec.params_cls()

Aircraft

Environment	Cite as	Observation	Action	Tracked	PID	MPC	Physics
plane	plane-v1	(10,)	(2,)	altitude (m)	yes	yes	contract
plane_energy	plane_energy-v1	(10,)	(2,)	altitude (m)	yes	yes	contract
plane_sine	plane_sine-v1	(10,)	(2,)	altitude (m)	yes	yes	contract
plane3d_heading	plane3d_heading-v1	(15,)	(3,)	altitude (m)	yes	yes	contract
plane3d_circle	plane3d_circle-v1	(17,)	(3,)	altitude (m)	yes	yes	contract
plane3d_racetrack	plane3d_racetrack-v1	(21,)	(3,)	altitude (m)	yes	yes	contract
plane3d_figure8	plane3d_figure8-v1	(19,)	(3,)	altitude (m)	yes	yes	contract
patrol	patrol-v1	(26,)	(3,)	slot error (m)	yes	yes	contract
patrol_bearing_only	patrol_bearing_only-v1	(21,)	(3,)	measured range (m)	yes	no	contract

patrol_bearing_only -- PID present -- a lead-state estimator feeding the same pursuit law the full-observation variant uses. Range with azimuth and elevation is a complete relative-position measurement, so the only genuinely unobservable quantity is the lead's HEADING, which the commanded slot needs because the slot is expressed in the lead's frame; it is recovered by differencing the estimated relative position and filtering. Measured performance matches the full-observation expert (4 of 8 seeds complete, ~229 m settled slot error vs ~260 m), so the partial observation costs essentially nothing here. No MPC, and the reason is the withheld observation rather than the manoeuvring lead. This note used to blame the lead, on the grounds that an MPC would need its future trajectory as a time-varying parameter. That holds for a CasADi model and not for a gradient planner: patrol now ships a GradientMPC that differentiates step_env, and because the lead is scripted and deterministic the plan propagates it for free. What blocks one here is that the planner reads the slot error out of the state, which is precisely what this variant withholds. Handing it the true state anyway would make it an oracle on a task defined by what is hidden, so it needs a planner built on the estimator.

Process Control

Environment	Cite as	Observation	Action	Tracked	PID	MPC	Physics
cstr	cstr-v1	(3,)	(1,)	C_a (mol/L)	yes	yes	contract
first_order	first_order-v1	(2,)	(1,)	x	yes	yes	contract
four_tank	four_tank-v1	(6,)	(2,)	h1 (m), h2 (m)	yes	yes	contract
ph_neutralization	ph_neutralization-v1	(3,)	(1,)	pH	yes	yes	contract
distillation	distillation-v1	(6,)	(2,)	yD (mole fraction)	yes	yes	contract

Industrial / Energy

Environment	Cite as	Observation	Action	Tracked	PID	MPC	Physics
glass_furnace	glass_furnace-v1	(5,)	(1,)	crown temperature (K)	yes	yes	contract
reactor	reactor-v2	(4,)	(1,)	neutron power (normalised)	yes	yes	contract
hvac	hvac-v1	(7,)	(1,)	zone air temperature (deg C)	yes	yes	contract
cement_kiln	cement_kiln-v1	(8,)	(2,)	discharge free lime (%)	yes	yes	contract
boiler_drum	boiler_drum-v1	(7,)	(2,)	drum level (m), drum pressure (bar)	yes	yes	contract

Renewable Energy

Environment	Cite as	Observation	Action	Tracked	PID	MPC	Physics
wind_turbine	wind_turbine-v1	(5,)	(2,)	electrical power (MW)	yes	yes	contract
battery	battery-v1	(5,)	(1,)	delivered power (MW)	yes	yes	contract