2026 Reading Log

First to use diffusion policies to hit (at the time) SOTA in robot policy learning — very cool, and first author Cheng Chi is a fellow Columbia alum (now founder of Sunday Robotics) :-)

Expands Diffusion Policy from 2D to 3D via point clouds. Feels like part of a broader trend in robotics — the shift from 2D to 3D data — that shows up well beyond diffusion policy (e.g. Waymo's LiDAR stack).

Brief skim, need a more in-depth pass. The original feed-forward 3D transformer for point maps — infers camera params, depth, and point clouds directly from images.

Sort of has memory? SOTA as of ~July 2026 — a lot of follow-up work has already built on this dual implicit memory idea (separate spatial vs. semantic representations).

High-level overview of the robotics-ML space — GROD (Gemini Robotics On-Device) and Gemini Robotics-ER (Embodied Reasoning) as two halves of the same family.

Builds on VGGT. A big open problem in these models is no absolute sense of scale — LingBot-Map tackles this with a persistent, linearly-scaling spatial memory instead of reprocessing everything each frame.

Brief skim — zero-shot prediction of CRISPRi knockdown responses in unseen cell lines. Planning on submitting to this one.

The first case I've seen of using RL instead of pre-training to bake in synthesizability constraints. Fits the general pattern: LLM/pre-training → fine-tune with RL.

The field is shifting toward 3D-equivariant/diffusion techniques (i.e. structure-based drug design), but Saturn shows SMILES/language-based models are still great at learning chemical properties. I'd bet on both: a hybrid of language-based/SMILES + diffusion.

In progress.

In progress.

In progress.

Written by my Probabilistic Models & ML professor, one of the leading experts in the field. Wrote my reading notes on this one!

Wrote my reading notes on this one too!

An alternative to Adam — on my to-read list. From what I gather, it orthogonalizes the update using the full 2D weight matrix, but I need to actually dig in.

Robotics!

On what a meaningful human life looks like when thinking is best done by machines.

Notes on LLMs.

Math notes.