ihdp-f16

IHDP Agent for F-16 Longitudinal Control (TensorAeroSpace)

This repository contains an Incremental Heuristic Dynamic Programming (IHDP) agent trained for longitudinal control of the F-16, implemented with TensorAeroSpace.

The agent tracks a step reference in angle of attack (alpha). It uses two neural networks (Actor and Critic) and an online incremental model to adapt features during rollout.

Model architecture and files

Saved artifacts (see this repository files):

• config.json: environment and agent configuration (IO, actor/critic/incremental settings, optional reference signal snapshot)
• actor.h5: Keras weights for the Actor network
• critic.h5: Keras weights for the Critic network
• incremental_model/:
  • F.npy, G.npy: incremental model matrices
  • delta_xt.npy, delta_ut.npy (if available): last gradients windows used during identification

How to use

import os
import numpy as np
import gymnasium as gym
from tensoraerospace.agent.ihdp.model import IHDPAgent

# Load pretrained IHDP agent (either from local folder or HF Hub repo ID)
repo_id = "<username>/<repo_name>"  # or path to a local saved folder
agent = IHDPAgent.from_pretrained(repo_id, access_token=os.getenv("HF_TOKEN"))

# Create F-16 longitudinal environment (must match config.json IO/params)
env = gym.make(
    "LinearLongitudinalF16-v0",
    number_time_steps=2002,
    initial_state=[[0], [0], [0]],
    reference_signal=np.zeros((1, 2002)),  # replace with your reference
    use_reward=False,
    state_space=["theta", "alpha", "q"],
    output_space=["theta", "alpha", "q"],
    control_space=["ele"],
    tracking_states=["alpha"],
)

# Example rollout (reference must be shaped [1, T])
obs, info = env.reset()
reference = env.unwrapped.reference_signal
for t in range(reference.shape[1] - 3):
    u = agent.predict(obs, reference, t)
    obs, r, terminated, truncated, info = env.step(np.array(u))
    if terminated or truncated:
        break

Reproduce the training/evaluation

This model card is based on the notebook example/general_examples/example_ihdp_beautiful.ipynb with the following settings:

• Simulation time: 20 s, dt: 0.01 s → 2002 steps
• Tracking variable: alpha (angle of attack)
• Reference: step of 5° (converted to radians)
• Initial state: [theta=0, alpha=0, q=0]

Results

From the evaluation run in the notebook:

• MAE (alpha): 0.042348 rad (2.426°)
• RMSE (alpha): 0.069442 rad (3.979°)
• Max error (alpha): 0.204428 rad (11.713°)
• Settling time (95%): 2.87 s

Control (elevator) statistics depend on limits and units used in the environment. Ensure your environment configuration matches config.json.

Intended use and limitations

• Intended for research and educational use on aircraft longitudinal control tasks.
• This is a simulation-trained controller; it is not validated for real-world flight.
• Performance is sensitive to environment configuration and reference signals. Always align your env settings with the provided config.

How this repository was created

Artifacts were saved using IHDPAgent.save(...)/save_pretrained(...) and uploaded with IHDPAgent.push_to_hub(...) in TensorAeroSpace.

Citation

If you use this work, please cite TensorAeroSpace.

@software{TensorAeroSpace,
  title        = {TensorAeroSpace: Open source deep learning framework for aerospace objects},
  author       = {Mazaev, Artemiy and Davydov, Vasily and Li, Yakov},
  year         = {2025},
  url          = {https://github.com/mr8bit/TensorAeroSpace},
}