bestak/uav-navigation-sasp

SASP: Saliency-Augmented Static Policy (Phase II)

About this Model

This model is a supplementary artifact for the paper "Saliency-Driven Active Perception for UAV Visual Re-Localization using Reinforcement Learning" (Vojtěch Bešťák, Václav Truhlařík, Libor Přeučil), based on the CTU FIT master's thesis of the same title (Vojtěch Bešťák, 2026).

It is trained with active-perception-uav-nav, a research framework for visual UAV navigation using RL/IL over aerial orthophoto and land-use grid environments.

Key	Value
Algorithm	PPO
Environment	aerial_grid
W&B group	exp2.2-aerial-ppo-ablation-sasp_3
Run ID	y8w84vhm
Reward weights	dist=1.096, view=19.445, time=-0.100, success=311.2
Max steps / episode	500
Frame stack	4
Seed	4
Python (training)	3.11.14
stable-baselines3	2.2.1
PyTorch	2.10.0
Gymnasium	0.29.1
Arena size	1000 m
Success radius	4 cells
Saliency channel	yes

Quantitative Performance

Results are pooled over 3200 evaluation arenas per seed, across five independently trained seeds, in continuous environments over authentic ČÚZK orthophotos. In-domain uses the five training regions; zero-shot uses two held-out regions with no geographic overlap with training.

Metric	In-domain	Zero-shot
Success Rate	99.42% ± 0.56	98.51% ± 1.63
Optimality Score	0.637 ± 0.012	0.602 ± 0.011
Path Extension	1.097 ± 0.030	1.088 ± 0.028
Average Localization Error	51.27 m ± 0.33	46.68 m ± 0.33
Recall@1	0.667 ± 0.001	0.700 ± 0.002

SASP closes 27.6% of the Naive–Oracle localization-error gap in-domain and 25.1% zero-shot, without retraining on the held-out regions.

Observation and Action Spaces

Key	Shape	dtype	Notes
camera	(4, 84, 84, 3)	uint8	Aerial orthophoto RGB crop, normalised to [0,1] inside extractor
visited_mask	(4, 84, 84)	uint8	Cells visited in the current episode crop
saliency	(4, 84, 84)	uint8	SelaVPR saliency crop
goal_info	(24,)	float32	Egocentric telemetry and navigation state

Action space: Box([0, -1], [1, 1], (2,), float32) — [velocity ∈ [0,1], angular_rate ∈ [-1,1]]

Standalone Inference (no active-perception-uav-nav package required)

Install the minimal deps (versions match the training environment):

pip install "stable-baselines3==2.2.1" "torch==2.10.0" "gymnasium==0.29.1" huggingface-hub

import importlib.util, sys, numpy as np
from huggingface_hub import hf_hub_download
from stable_baselines3 import PPO

REPO_ID = "bestak/uav-navigation-sasp"

# 1. Download and load the feature extractor (pure PyTorch, no repo import needed)
fe_path = hf_hub_download(REPO_ID, "feature_extractor.py")
spec = importlib.util.spec_from_file_location("_fe", fe_path)
mod  = importlib.util.module_from_spec(spec)
sys.modules["_fe"] = mod
spec.loader.exec_module(mod)

# 1b. Stub out drone_navigation so cloudpickle can resolve ALL saved class references
#     (lr_schedule, policy_kwargs, etc.) without the package being installed.
import types as _types
for _name in ["drone_navigation", "drone_navigation.models",
              "drone_navigation.models.feature_extractor_aerial",
              "drone_navigation.models.feature_extractor_landuse"]:
    sys.modules.setdefault(_name, _types.ModuleType(_name))
sys.modules["drone_navigation.models.feature_extractor_aerial"].AerialFeaturesExtractor = mod.AerialFeaturesExtractor

# 2. Load the model -- inject the extractor class so cloudpickle can resolve it
model = PPO.load(
    hf_hub_download(REPO_ID, "best_model.zip"),
    custom_objects={
        "features_extractor_class": mod.AerialFeaturesExtractor,
    },
    device="cpu",
)

# 3. Run a single forward pass with a dummy observation
obs = {
    "camera":       np.zeros((4, 84, 84, 3), dtype=np.uint8),
    "visited_mask": np.zeros((4, 84, 84),    dtype=np.uint8),
    "goal_info":    np.zeros(24, dtype=np.float32),
    "saliency":     np.zeros((4, 84, 84),    dtype=np.uint8),
}
action, _ = model.predict(obs, deterministic=True)
print("Action:", action)

Note: custom_objects overrides the cloudpickled class reference, which is why the active-perception-uav-nav package is not required for loading. See inference.py in this repo for the full example including environment rollouts.

Full Inference with the active-perception-uav-nav Package

pip install git+https://github.com/bestak/active-perception-uav-nav.git
# also requires map data -- see the repo README for data preparation

from huggingface_hub import hf_hub_download
from stable_baselines3 import PPO
from drone_navigation.config.experiment_config import ExperimentConfig
from drone_navigation.envs.factory import create_env

REPO_ID = "bestak/uav-navigation-sasp"

# When drone_navigation is installed, the extractor class resolves automatically
model = PPO.load(hf_hub_download(REPO_ID, "best_model.zip"), device="cpu")

cfg = ExperimentConfig.from_json(hf_hub_download(REPO_ID, "config.json"))
cfg.n_envs = 1
env = create_env(cfg)

obs, _ = env.reset()
for _ in range(cfg.max_steps):
    action, _ = model.predict(obs, deterministic=True)
    obs, reward, terminated, truncated, info = env.step(action)
    if terminated or truncated:
        print("Episode done. Target reached:", info.get("is_target_reached"))
        break

env.close()

See inference.py in this repo for a more complete example with multi-episode evaluation.

Training

git clone https://github.com/bestak/active-perception-uav-nav.git
cd active-perception-uav-nav
uv sync
drone-train-rl --env_type aerial_grid ...

Citation

If you use this model, please cite the paper:

@misc{bestak2026:saliencyactiveperception,
  author = {Be{\v{s}}{\v{t}}{\'a}k, Vojt{\v{e}}ch and Truhla{\v{r}}{\'i}k, V{\'a}clav and P{\v{r}}eu{\v{c}}il, Libor},
  title  = {Saliency-Driven Active Perception for UAV Visual Re-Localization using Reinforcement Learning},
  year   = {2026}
}