iiinova — where your digital self keeps evolving

Joe Watson

@joemwatson.bsky.social

Postdoctoral researcher in robotics & machine learning for control (A2I, Oxford Robotics Institute) previously PhD @ TU Darmstadt / DFKI, Google DeepMind robotics intern, CMR Surgical, University of Cambridge joemwatson.github.io

Posts Replies Media

Reposted by Joe Watson

Daniel Palenicek @daniel-palenicek.bsky.social · 22/04/2026

Headed to Rio for #ICLR 🇧🇷 come say hi at our poster! XQC: a principled look at critic optimization. BN+WN+cross-entropy → condition numbers orders of magnitude smaller than baselines. SOTA on 70 continuous ctrl tasks w/ 4.5× less params. 📅 Thu, 1030–1300 📍 Pavilion 4, #4518

031

Joe Watson @joemwatson.bsky.social · 13/02/2025

Interesting, thanks! From what I remember the model rollout horizon in MBPO was often only 1 timestep (or started at 1 timestep and was extended during learning)

130

Joe Watson @joemwatson.bsky.social · 13/02/2025

This sounds a lot like MBPO, is it essentially MBPO-style augmentation on top of TD-MPC?

110

Joe Watson @joemwatson.bsky.social · 02/12/2024

I think I read Phillip Pullman's His Dark Materials trilogy after the Potter books. If you don't know the series, it's an interesting blend of fantasy, science fiction, and religious themes

080