Aevanis-Kinematics: Non-Rigid Contact Physics & Anti-Identity-Swap Engine

Sub-4ms Physical Disentanglement & Grasp Boundary Head for Robotic Manipulation.

Developed by Aevanis AI Lab, Aevanis-Kinematics resolves non-rigid surface deformations, tactile interaction vectors, and prevents visual identity swapping during high-speed robotic manipulation.

---

The Identity-Swap & Contact Boundary Problem

During fast robotic grasping, picking, and dexterous manipulation, conventional computer vision models fail when two objects touch:

• Object Merging: Gripper and target object merge into a single visual contour.
• Identity Swapping: Occlusions cause trackers to interchange object IDs.
• Lack of Contact Force Awareness: Standard bounding boxes cannot infer whether physical contact has occurred.

Aevanis-Kinematics solves this at 100+ FPS by predicting:

1. Dense 14x14 Contact Boundary Segmentation: Segments the high-strain boundary where two physical bodies interface.
2. Relative Motion Flow Vectors (dX, dY): Tracks dynamic slip and shearing trajectories.
3. Deformation & Contact Pressure Score: Regresses normal contact force intensity in [0, 1].

---

Available Model Weights

• aevanis_kinematics_nano.pt (0.42 MB): Built on Av-Vision-Nano (7.68M).
• aevanis_kinematics_core.pt (0.62 MB): Built on Av-Vision-Core (24.35M). Enhanced dexterous contact sensitivity.

---

Verified Performance & Loss Metrics

• Contact Boundary BCE Loss: 0.1484
• Motion Flow Direction MSE: 0.2144
• Contact Deformation / Pressure MSE: 0.0807
• Execution Latency: < 3.2 ms

---

Quickstart

import torch
from model import AevanisKinematicsAgent

agent = AevanisKinematicsAgent(model_scale="nano")
agent.load_head("aevanis_kinematics_nano.pt")
agent.eval()

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

print("Contact Boundary Map:", kin["contact_map"].shape)
print("Relative Motion Flow :", kin["motion_flow"])
print("Contact Pressure     :", kin["contact_pressure"])

---

License

Apache-2.0.