Alexandre Brown 🇨🇦

🚀 I'm excited to share our new paper: SegDAC: Segmentation-Driven Actor-Critic for Visual Reinforcement Learning 🧠 SegDAC combines large vision models with online RL to reason about its environment at the object and sub-object level, avoiding noisy pixel-level reasoning. 🛠️ Using YOLO-World and SAM, SegDAC breaks the scene into semantically meaningful segments and learns to attend to a variable number of segments and proprioceptive signals, focusing on the most relevant information to complete the task. ⚡ Trained purely with online RL, without human labels or demonstrations. 🏆 Outperforms previous online RL state-of-the-art methods across all difficulty levels on our challenging visual generalization benchmark, with up to 2x better visual generalization in the hardest setting. 📄 Paper: arxiv.org/pdf/2508.09325 🌐 Project Page: segdac.github.io Work done with @GlenBerseth at @Mila_Quebec #ReinforcementLearning #RobotLearning #ArtificialIntelligence #Robotics

What a lovely way to continue this debate in 2026 💪

🚀 I vibecoded yet another autoresearch tool This one works with SLURM clusters and lets Claude Code agents run experiments completely hands-off for weeks. It's called xgenius — open source, works with ANY codebase. github.com/roger-creus/xg…

Unveiling our new startup Advanced Machine Intelligence (AMI Labs). We just completed our seed round: $1.03B / 890M€, one the largest seeds ever, probably the largest for a European company. We're hiring! [the background image is the Veil Nebula - a picture I took from my backyard, most appropriate for an unveiling] More details here: techcrunch.com/2026/03/09/yan…

@francoisfleuret Do we have examples of hypernetworks working at large scale? That feels like the missing piece here.

BTW are hyper-networks a thing of the past?

I will be giving a talk at UPenn @GRASPlab tomorrow 3-4PM EST on my research in sim/robotics. I’ll be discussing how sim integrated robot learning can drive and accelerate robotics progress. If you are in the area let’s meet up! Link with more details in the thread

We came away pleasantly surprised by how tractable these models are to decode, how different video encoders feel from language (many techniques won’t transfer), and how brain-like these representations look, reviving old “petri dish” conversations about studying the brain.

@N8Programs This feels wrong but is beautiful at the same time.

Beat it by having Codex hand-craft weights: gist.github.com/N8python/02e41… 100% accuracy on 10 million random test cases w/ only 343 parameters. As a bonus, it uses the vanilla Qwen3 architecture, just with the right weights.

I spent $0 and a weekend vibecoding an @openclaw setup that I text to run experiments for me. In the process, I ended up with bespoke software for self-managing a personal cluster. Also, it now comes up with its own experiments if I don’t have enough running. Blog post link ⬇️

@F_Vaggi @rasbt @rryssf_ You deserve my follow for this ✅

@rasbt @rryssf_ I follow a pretty rigid rule where if someone who doesn't work in ML in a technical role tweets hyped up stuff that's clearly designed to go viral to build up clout, I mute them.

DeepMind just did the unthinkable. They built an AI that doesn't need RAG and it has perfect memory of everything it's ever read. It's called Recursive Language Models, and it might mark the death of traditional context windows forever. Here's how it works (and why it matters way more than it sounds) ↓

4/ Results & What Worked Delay Target performed best with the most stable learning curve, followed by Rainbow DQN and SAC which showed comparable performance. PPO significantly underperformed. The pattern is clear: off-policy Q-learning methods dominated. When you can't parallelize environments and data arrives at real-time speed, the ability to reuse past experiences via replay buffers is critical.

@Yuchenj_UW You can "read" paper faster now but do you still understand a paper as deeply if you don't read it yourself, that's the question.

Tbh, if I had Claude Code, Gemini, and ChatGPT during my PhD, I’d probably have graduated in 1 year instead of 5.5 years. My PhD was ~50% coding, 25% writing/polishing my papers, 25% reading others' papers. AI now accelerates each by at least 10×. Nothing will ever be the same.

@Stone_Tao @bharthulwar Very insightful, I wonder if there's a connection between this and how Off-policy methods get this for free with a Replay Buffer.

GPU parallelized envs have accelerated RL, but most implementations exhibit critical instability when running on-policy RL with short rollouts. We present Staggered Environment Resets. A few lines of code are all you need! Presenting today, 4:30PM poster 310 #NeurIPS2025 🧵(1/8)

@creus_roger Well-deserved Roger!

Accepted to NeurIPS 2025 as a spotlight (top ~3% of submissions)!!! ✨🌟 Paper: arxiv.org/abs/2506.15544 Code: github.com/roger-creus/st…

@ebanks_keyshawn @arnie_hacker Can you link a paper that supports your claim ?

@arnie_hacker from what i’ve seen anything imitation + RL. For tasks like this you might want to throw an LLM in the loop

Crazy to me how far away we are from generalizable robotics. Nvidia Gr00t N1.5, a SOTA model: - trained on 1K H100 GPUs - with data augmentation - and future latents alignment (& human data) - and 1K teleop demos Drops success rate by ~14% when there are new objects in the observation. How will the humanoids react when they're deployed in real worlds with humans walking around for instance (haven't seen any in the datasets so far)

Everyone uses DAgger in robot learning, but most papers barely mention how they do it … 😕 We’ve found a few subtle details that make a big difference (which differ from the standard practice today):💡 - How to collect corrections? Standard human “take-over” corrections create discontinuities. A compliant interface that gently nudges the on-policy roll-out works much better. - How to update the policy? Instead of standard finetuning, we found that learning residual networks works better — more stable and flexible, and lets you plug in new modalities (e.g., force on top of a position-only base policy). Our paper Compliant Residual DAgger summarizes these and other interesting findings in detail 👇

I do not think you can pursue meaningful research without (1) some grandiose delusion about your abilities (2) a sense of esthetics and harmony to judge ideas still free of experimental confirmation (3) an unreasonable taste for the required tangible work (e.g. programming)

> Be AI PhD student > Submit paper to conference > LLM slop reviews > Rejected > Concurrent paper with same method accepted > Resubmit to next conference > Reviewer points to concurrent paper which was accepted by last conference > Lack of novelty > Rejected

People are giving up on AI conferences due to non-sensical / non-professional reviews. It's time to stop this madness.

@ScholarshipfPhd Hi

Say hi and I’ll recommend a research topic that perfectly fits your profile.