ImprovedLAPANEnv -- LAPAN LSU-05 NG pitch tracking (normalized, RL-friendly)¶

This page documents the ImprovedLAPANEnv environment used for training and evaluating reinforcement learning controllers for the longitudinal channel of the LAPAN Surveillance Aircraft (LSU)-05 NG. The environment provides a normalized observation and action interface, a shaped reward for precise and smooth pitch tracking, and realistic termination conditions.

The reference implementation lives in tensoraerospace/envs/lapan.py (class ImprovedLAPANEnv).

Summary¶

Observation space: 4D, normalized to [-1, 1]
Action space: 1D, normalized to [-1, 1] (elevator deflection command)
Goal: track a time-varying pitch reference while keeping control smooth and economical
Reward: quadratic costs on pitch tracking error, pitch-rate mismatch to reference dynamics, absolute input, input rate and input jerk (with scaling)
Termination: exceeding safe pitch envelope or reaching time horizon

Observation and Action¶

Let \(\theta\) be the current pitch angle (rad), \(q\) the pitch rate (rad/s), and \(\theta_{ref}(t)\) the target pitch.

The observation at step \(t\) is a 4-vector:

\[ \mathbf{o}_t = \big[\underbrace{\mathrm{clip}\!\Big(\frac{\theta_{ref}(t) - \theta(t)}{\theta_{max}},\,-1,\,1\Big)}_{\text{[0] norm\_pitch\_error}},\; \underbrace{\mathrm{clip}\!\Big(\frac{q(t)}{q_{max}},\,-1,\,1\Big)}_{\text{[1] norm\_q}},\; \underbrace{\mathrm{clip}\!\Big(\frac{\theta(t)}{\theta_{max}},\,-1,\,1\Big)}_{\text{[2] norm\_\(\theta\)}},\; \underbrace{u_{t-1}}_{\text{[3] prev action}}\big] \]

where

\(\theta_{max} = 20°\) (converted to rad internally),
\(q_{max} = 5°/\text{s}\) (converted to rad/s internally),
\(u_{t-1} \in [-1, 1]\) is the previously applied normalized elevator command.

Important: The observation vector is already normalized and ready to use. Index [0] contains the normalized pitch tracking error, which is directly usable for proportional control.

The action is a single normalized command \(u_t \in [-1, 1]\). It maps to a physical elevator deflection (rad):

\[ \delta_e(t) = u_t \cdot \delta_{e,max}, \quad \delta_{e,max} = 25° \text{ (in rad)}, \]

which is passed directly to the underlying LAPAN model in radians.

Reward function¶

The environment uses a shaped reward with five terms promoting accuracy and smoothness while penalizing excessive actuation. The per-step reward is

\[ r_t = -\,s\,\Big(\,w_\theta\,e_\theta^2\; +\; w_q\,e_q^2\; +\; w_u\,u_t^2\; +\; w_{\Delta}\,(\Delta u_t)^2\; +\; w_{\Delta^2}\,(\Delta^2 u_t)^2\Big), \]

with the default weights and scale:

\[w_\theta=5.0,\quad w_q=0.2,\quad w_u=0.003,\quad w_\Delta=0.01,\quad w_{\Delta^2}=0.001,\quad s=0.1.\]

The error terms are defined as

\[ e_\theta = \frac{\theta(t)-\theta_{ref}(t)}{\theta_{max}},\qquad e_q = \frac{q(t)-\dot\theta_{ref}(t)}{q_{max}}, \]

where \(\dot\theta_{ref}(t)\) is a finite-difference derivative of the reference pitch computed with the simulation time step \(\Delta t\). The input smoothness terms use

\[ \Delta u_t = u_t - u_{t-1},\qquad \Delta^2 u_t = u_t - 2 u_{t-1} + u_{t-2}. \]

Notes:

A larger \(w_\theta\) emphasizes precise pitch tracking.
\(w_q\) damps response relative to the reference slope, reducing overshoot/oscillation.
\(w_u, w_\Delta, w_{\Delta^2}\) regularize energy usage and command smoothness, suppressing chattering.
The overall scale \(s\) keeps rewards in a compact numerical range for RL stability.

Termination and truncation¶

Safety termination: if \(|\theta| > \theta_{max}\), the episode terminates early and a large penalty (\(-100\)) is applied at that step.
Truncation: the episode truncates when the configured horizon (number of time steps) is reached.

Episode dynamics¶

At each step:

The agent outputs \(u_t \in [-1, 1]\).
The environment clamps \(u_t\) to \([-1, 1]\), maps it to \(\delta_e\) in radians, and advances the internal LAPAN model.
The observation \(\mathbf{o}_{t+1}\) is built using normalized signals.
The reward \(r_t\) is computed as above.
Termination/truncation is checked.

Usage example¶

import numpy as np
from tensoraerospace.envs.lapan import ImprovedLAPANEnv
from tensoraerospace.signals.standard import sinusoid_vertical_shift
from tensoraerospace.utils import generate_time_period, convert_tp_to_sec_tp

dt = 0.01
tn = 200
tp = generate_time_period(tn=tn, dt=dt)
tps = convert_tp_to_sec_tp(tp, dt=dt)
number_time_steps = len(tp)

# reference: smooth sinusoidal pitch in radians (amplitude 1 deg)
reference_signal = np.reshape(
    sinusoid_vertical_shift(
        tp=np.asarray(tps), frequency=0.05, amplitude=np.deg2rad(1.0), vertical_shift=0.0
    ),
    (1, -1),
)

# Initial state: [u, w, q, theta] -- depends on LAPAN model state order
initial_state = np.array([0, 0, 0, 0], dtype=np.float32)

env = ImprovedLAPANEnv(
    initial_state=initial_state,
    reference_signal=reference_signal,
    number_time_steps=number_time_steps,
    initial_elevator_deg=0.0,
    use_initial_action_on_first_step=True,
    dt=dt,
)

obs, info = env.reset()
done = False
while not done:
    # simple proportional control on normalized pitch error in [-1, 1]
    u = float(np.clip(2.0 * float(obs[0]), -1.0, 1.0))
    obs, reward, terminated, truncated, info = env.step(np.array([u], dtype=np.float32))
    done = bool(terminated or truncated)

Implementation hints¶

Observation normalization bounds (\(\theta_{max}, q_{max}\)) and elevator limits are defined inside the class and can be tuned if needed.
Reward weights are exposed as instance attributes (w_pitch, w_q, w_action, w_smooth, w_jerk, reward_scale) and can be changed to match specific control objectives.
The LAPAN model state order is [u, w, q, theta]. The environment extracts \(q\) at index 2 and \(\theta\) at index 3. Observations are normalized to [-1, 1].
The LAPAN model expects elevator input in radians; the environment handles the conversion from the normalized action space internally.

References¶

tensoraerospace/envs/lapan.py -- full environment implementation
LAPAN LSU-05 NG dynamics model: tensoraerospace/aerospacemodel/lapan.py