astralhf/astral-drone-models

Astral Drone Models

Pre-trained inference models from Astral — the open-source autonomous drone platform. These models are used in the Astral drone daemon running on Jetson Orin Nano hardware (M1-A quadcopter, M1-G ground rover).

All models are ONNX or GGUF format, ready for on-device inference without a training environment.

---

Models

yolov8n_domain_v3.onnx — Aerial Domain Detector

Size: 11.7 MB · Format: ONNX · Architecture: YOLOv8n

9-class aerial domain detector trained on 48,000 sim+real images across three training rounds. Classifies the visual domain of an aerial frame (indoor corridor, outdoor open, urban rooftop, etc.) to route perception appropriately.

• Input: [1, 3, 640, 640] float32 RGB
• Output: standard YOLOv8 detection head
• mAP50 after round 3: 0.384 (drone AP50: 0.087)
• Training detail: blog post · technical report

---

vlm_lora_v1_q4km.gguf — VLM LoRA (Q4_K_M)

Size: 1.8 GB · Format: GGUF Q4_K_M · Base: Qwen2.5-VL-3B

Vision-language model fine-tuned with LoRA on drone-perspective navigation data. Used as the semantic target selector in Astral's modular autonomy stack — it answers "which object in this frame is the navigation goal?" and hands off to the metric depth module for 3D localization.

Run with llama.cpp or any GGUF-compatible runtime. Requires the paired mmproj file below.

• Pair with: vlm_lora_v1_mmproj.gguf
• Role in stack: semantic target selection (not end-to-end control)
• Architecture detail: Closing the Metric Gap

---

vlm_lora_v1_mmproj.gguf — VLM Multimodal Projector

Size: 1.2 GB · Format: GGUF · Base: Qwen2.5-VL-3B mmproj

Multimodal projector for vlm_lora_v1_q4km.gguf. Required — the main GGUF will not load without it.

Pass to llama.cpp with --mmproj vlm_lora_v1_mmproj.gguf.

---

policy_v1.onnx + policy_v1.onnx.data — Reactive Policy

Size: 121 KB total · Format: ONNX (external data)

Lightweight MLP reactive policy for collision avoidance and low-level flight stabilization. Runs alongside the VLM planner — the VLM sets the goal, the reactive policy handles moment-to-moment obstacle response.

Both files must be present in the same directory; policy_v1.onnx.data contains the weight tensors.

• Input: state vector (normalize with policy_v1_state_norm.npy before inference)
• Output: velocity command delta

---

policy_v1_state_norm.npy — State Normalization

Size: 224 B · Format: NumPy array

Mean/std normalization constants for policy_v1.onnx input. Load with numpy.load('policy_v1_state_norm.npy', allow_pickle=True).item() — returns {'mean': ..., 'std': ...}.

---

depth_v1.onnx — Monocular Depth Model

Size: 1.6 MB · Format: ONNX

Lightweight monocular depth estimation model for metric grounding. Converts the VLM's object identification into a 3D position estimate for the planner. Designed to run in real time on Jetson Orin Nano.

• Input: [1, 3, H, W] float32 RGB (model handles resize)
• Output: [1, 1, H, W] relative depth map

---

Usage

The Astral daemon (daemon.py) downloads these automatically via setup_models.py:

python3 setup_models.py --domain-detector --reactive-policy --depth-model --vlm-drone

Or download individually with huggingface_hub:

from huggingface_hub import hf_hub_download

# Domain detector
hf_hub_download("astralhf/astral-drone-models", "yolov8n_domain_v3.onnx")

# VLM (both files required)
hf_hub_download("astralhf/astral-drone-models", "vlm_lora_v1_q4km.gguf")
hf_hub_download("astralhf/astral-drone-models", "vlm_lora_v1_mmproj.gguf")

# Reactive policy (both files + norm required)
hf_hub_download("astralhf/astral-drone-models", "policy_v1.onnx")
hf_hub_download("astralhf/astral-drone-models", "policy_v1.onnx.data")
hf_hub_download("astralhf/astral-drone-models", "policy_v1_state_norm.npy")

# Depth
hf_hub_download("astralhf/astral-drone-models", "depth_v1.onnx")

Architecture

These models implement the separation principle: semantic understanding (VLM) is decoupled from metric geometry (depth) and low-level control (reactive policy). This architecture is why end-to-end VLMs fail drone navigation while the modular stack succeeds — detailed in Closing the Metric Gap.

Related

• Astral SDK
• Yonder dataset — 4.65M-frame drone navigation dataset used to develop these models
• Research hub
• Blog

License

Models are released under CC-BY-NC-4.0. Research and non-commercial use permitted with attribution. Contact astral.us for commercial licensing.