Aevanis/Aevanis-Vision-Nano

Aevanis Vision Nano (Universal Visual Eye)

High-Throughput Foundation Visual Backbone for Robotics, Autonomous Navigation, and Real-Time Spatial Perception.

Developed by Aevanis AI Lab, Av-Vision-Nano is an edge-native (7.68M parameter) hybrid vision transformer designed specifically for ultra-high FPS inference on edge hardware (Jetson, Raspberry Pi, embedded systems) and high-throughput server streams.

---

Architectural Highlights

• 750 to 900 FPS throughput on Server GPUs (Tesla T4 / RTX 4090).
• 180 to 240 FPS real-time control rate on NVIDIA Jetson Orin Nano.
• 15 to 25 FPS perception on Raspberry Pi 5.
• Sub-Pixel Geometric Stem: ConvNeXt-inspired 4x downsampling preserving high-frequency edge gradients before transformer tokenization.
• 196 Spatial Disparity Tokens: Extracts a 14x14 grid of continuous spatial tokens encoding metric depth and surface boundaries without physical laser LiDAR sensors.

---

Hardware Latency & Throughput Benchmark

Hardware Target	Batch Size	Latency (ms)	Inference Rate (FPS / Hz)	Deployment Target
NVIDIA RTX 4090	1	1.1 ms	900+ FPS	Multi-Agent Server Simulation
Tesla T4 (fp16)	1	1.3 ms	750 FPS	Cloud Perception Hub
Jetson Orin Nano	1	4.2 ms	240 FPS	Agile Quadrupeds / Drones
Desktop CPU (i7)	1	22.0 ms	45 FPS	Local Real-Time Vision
Desktop CPU (INT8 Quantized)	1	12.5 ms	80 FPS	Low-Latency CPU Deployment
Raspberry Pi 5	1	40.0 ms	25 FPS	Low-Cost Autonomous Rovers

---

Pre-Training Dataset

Trained on the Aevanis Continuous Egocentric & High-FPS Dynamics Dataset:

• 166,440 continuous, temporally synchronized video frames across 67 diverse high-motion sequences (48.2 GB raw temporal footage).
• Captures continuous physical dynamics, non-rigid kinematics, complex spatial occlusions, continuous camera ego-motion, and desktop GUI interaction streams.
• Pre-trained for 20 deep epochs with 75% patch masking and Normalized Sobel Boundary Gradient loss to internalize surface geometry and spatial continuity.
• Zero static web scraping. 100% continuous temporal stream pre-training.

---

Quickstart (3-Line Inference)

import torch
from model import AevanisVisionBackbone

# 1. Load Universal Eye Backbone
model = AevanisVisionBackbone.from_pretrained("pytorch_model.bin", model_scale="nano")

# 2. Extract 14x14 Spatial Features
img = torch.randn(1, 3, 224, 224)
with torch.no_grad():
    features = model(img)

# Shape: [1, 196, 256] -> 196 physical spatial tokens ready for downstream heads
spatial_tokens = features["spatial_tokens"]

---

License

Apache-2.0. Open for commercial, research, and edge robotics integration.