Aevanis-Traverse: Monocular Pseudo-LiDAR & Continuous Foothold Engine

High-Speed 150+ FPS Terrain Cognition Head for Quadruped, Bipedal, and Rover Robotics.

Aevanis-Traverse maps single monocular camera feeds into dense 3D spatial occupancy, foothold safety grids, terrain slope angles, and surface slip hazards without requiring heavy hardware LiDAR sensors.

---

Monocular 3D Pseudo-LiDAR Principle

Mechanical laser LiDAR units add physical weight, consume battery, and cost thousands of dollars. Aevanis-Traverse solves spatial geometry directly from camera photons:

1. Sub-Pixel Spatial Tokenization: The hybrid ConvNeXt stem preserves 2D high-frequency depth discontinuities and shadow gradients.
2. Dense 28x28 Transposed Upsampling: Expands the 14x14 latent tokens into a dense 28x28 traversability grid.
3. Multi-Task Geometric Affordance: Concurrently predicts ground normal vectors (pitch and roll slopes) and dynamic friction coefficients (slip risk) in a single sub-3ms forward pass.

---

Available Model Weights

• aevanis_traverse_nano.pt (2.52 MB): Built on Av-Vision-Nano (7.68M). 150+ FPS control loop.
• aevanis_traverse_core.pt (3.70 MB): Built on Av-Vision-Core (24.35M). Enhanced multi-terrain precision.

---

Verified Performance

• Dense Foothold BCE Loss: 0.1405
• Ground Slope Error: +-1.8 deg (Sub-2-degree precision for incline navigation)
• Surface Slip Hazard MSE: 0.00753 (Reaches the < 0.010 target threshold)
• Latency: < 2.8 ms (150+ Hz real-time robotics balance loop)

---

Quickstart

import torch
from model import AevanisTraverseAgent

agent = AevanisTraverseAgent(model_scale="nano")
agent.load_head("aevanis_traverse_nano.pt")

tokens = torch.randn(1, 196, 256)
with torch.no_grad():
    nav = agent(tokens)

foothold = nav["foothold_map"] # 28x28 safety grid
slope = nav["slope_angles"]    # [pitch, roll] in degrees
slip = nav["slip_risk"]        # [0 = firm ground, 1 = hazardous surface]

---

License

Apache-2.0.