HtW(s) Hack the World(s) Hackathon World Models · JEPA UTC Team · 19–20 June 2026

PCL‑JEPA · a view-independent 3D encoder

A joint‑embedding predictive encoder whose representation stays put while the object rotates.

This playground compares two point‑cloud encoders on the same rotating object: a rotation‑invariant PCL‑JEPA embedding against a regular PointNet.

We believe a view‑independent 3D encoder is one of the key challenges to overcome for the scene understanding of embodied agents equipped with 3D sensors.

Built 19–20 June 2026 Where ESIEE Paris · Noisy‑le‑Grand By Alexis Deseure--Charron · Robert Antaluca · Fadi Shafo · Yazan El-Mahmoud

object · —+ center frame

◆ rotation-invariant PCL-JEPA no file

PCA(2) · drag to orbit (3D)

—cos(embₜ, emb₀)

—drift

—emb dim

—var expl.

◇ regular PointNet no file

PCA(2) · drag to orbit (3D)

—cos(embₜ, emb₀)

—drift

—emb dim

—var expl.

angle 0°

embedding zoom 1.0×

Results · a view-independent embedding emerges

We observe far less drift / movement of an object's embedding as it rotates: a natural, view-independent embedding of the objects is produced.

Beyond per-object stability, a natural clustering emerges in the embedding space different shapes settle into distinct, well-separated regions while every rotation of a given shape stays put within its own cluster.