Michael Drolet

Interested in discrete latents? I’ll be presenting a new method for training discrete VAEs at ICLR in Rio 🇧🇷 📍 Pavilion 3, P3-605 🗓 Thu, Apr 23 @ 11:15 More details below 👇

Correction: Poster session is at 3:15 🙂

Interested in discrete latents? I’ll be presenting a new method for training discrete VAEs at ICLR in Rio 🇧🇷 📍 Pavilion 3, P3-605 🗓 Thu, Apr 23 @ 11:15 More details below 👇

Huge thanks to my co-authors: @firasalhafez, @aditya_bhatt, @Jan_R_Peters, and @OlegArenz. Grateful for the support from TU Darmstadt, DFG, and Hessian AI. If you'll be at ICLR (Rio 🇧🇷), come say hi! Questions & feedback welcome 🙏 @ias_tudarmstadt #ICLR2026

Step 3: Optimal target distribution. Solve a KL-constrained optimization problem to get a closed-form non-parametric target q*. Update the parametric encoder toward q*.

DAPS: Discrete Variational Autoencoding via Policy Search (accepted @iclr_conf) A principled way to train discrete autoregressive encoders — without straight-through gradients. Entropy reg. + KL trust region for stable training and compact latents. 🔗 drolet.io/daps

Proud to share our latest work, accepted at @iclr_conf 2026: APPLE! 🍎 TL;DR: APPLE is a novel reinforcement learning framework for solving active perception problems. #ICLR2026 #Robotics #MachineLearning #ActivePerception #RL @ias_tudarmstadt

🧵 Accepted at @iclr_conf ! Target networks stabilize bootstrapping in RL 🛡️ But induce slow-moving targets 🐢 Online networks adapt fast ⚡ But can diverge with function approximation 💥 𝗠𝗜𝗡𝗧𝗢🌿 uses the online network 𝗼𝗻𝗹𝘆 𝗶𝗳 𝗶𝘁 𝗰𝗮𝗻 — yielding faster and more stable RL. Here’s how 👇

@davide_tateo @liu_puze Thank you Davide for your mentorship. You will be missed!

My life chapter in Darmstadt is coming to an end. Today, I got a goodbye present from my dear friend (and amazing researcher) @liu_puze

@hbouammar I had a similar thought when I came across this paper recently.. arxiv.org/pdf/2205.11275

I read this paper in detail, and I am very sad! They literally re-do the optimal reward baseline work that we have known since forever, without even crediting the true authors in their derivations. The third screenshot is taken from: ieeexplore.ieee.org/stamp/stamp.js… As you see, they are almost identical! This is so wrong!! Stop this behavior @Microsoft ... sounds novel doesn't it ? Ridiculous to say the least - it has been done DECADES AGO! #AI #MachineLearning

🚀 We’re releasing #LocoMuJoCo2 — An extensive library for whole-body imitation & RL. 🔥 Highlights: - Supports MuJoCo & MJX - Clean JAX impl of PPO, GAIL, AMP, DeepMimic - 22K+ trajectories per humanoid - Add your humanoid easily 🔗 github.com/robfiras/loco-… 🧵Details 👇

@davide_tateo @Theo_Vincent_ @nabokovtimothy the best framework 𝘣𝘦𝘩𝘪𝘯𝘥 𝘫𝘢𝘹

@Theo_Vincent_ @nabokovtimothy PyTorch is the best framework 😉

Really cool and clean implementation of i-DQN in PyTorch from @nabokovtimothy✨ x.com/nabokovtimothy…

Something exciting just arrived at our lab! Here's a hint: precision mechanics, intricate wiring, and a touch of power. What could it be? A) High-end lab equipment B) Experimental space probe component C) Something that might walk past you in the hallway. Share your guesses!

@asurobot @davide_tateo @_bbelousov @SimonStepputtis Will always have a special place in my heart for IRL... Similar research, just in a different country 🙂 We’ve got some cool humanoids on the way—exciting times!!

@davide_tateo @_bbelousov @SimonStepputtis We already miss him, but I really can't wait to see what cool stuff his going to be doing at IAS! At least it's all in the family ;)

that looks better than the Figure demo

@davide_tateo @_bbelousov @Boris haha. it's my first day with x.. rookie mistake

@mdrolet01 @_bbelousov @Boris And that Boris is not @_bbelousov 😄