📣 Call for Blog Posts at #ICLR2026 @iclr_conf Following the success of the past iterations, we are opening the Call for Blog Posts 2026! iclr-blogposts.github.io/2026/about/#... Please retweet!

A practical introduction to (deep) RL, providing intuitions to understand the more recent algorithms. The plan is to start from tabular Q-learning and work our way up to Deep Q-learning (DQN). In a following post, I will continue on to Soft Actor-Critic (SAC) and its extensions.

Next Saturday, 𝗔𝗻𝘁𝗼𝗶𝗻𝗲 𝗣𝗶𝗿𝗿𝗼𝗻𝗲 will present Pollen Robotics & Hugging Face's open-source robots, including Reachy Mini, the SO-100 arm, the Amazing Hand and the Open Duck Mini. He will discuss the sim2real challenges of making the Open Duck Mini walk, and how […] [Original post on fosstodon.org]

RL102: From Tabular Q-Learning to Deep Q-Learning (DQN) - A Practical Introduction to (Deep) Reinforcement Learning araffin.github.io/post/rl102/

Package building with Pixi is being rolled out! Dive into our latest blog post on crafting C++ packages. And guess what? It’s not just for C++; Pixi plays nice with Python, Rust, ROS, Mojo, and beyond! prefix.dev/blog/pixi-b...

bash tricks permalink: wizardzines.com/comics/bash-... from our zine "Bite Size Command Line": wizardzines.com/zines/bite-s...

This is absolutely true -- this is a superb and much-needed consolidation of so much of modern RL. Kevin, inquiring minds want to understand the process you use to put this artwork together! @sirbayes.bsky.social Perhaps this is also the ultimate benchmark for Gemini Deep Research reports. ;-p

Weekend project: building a (site) search engine www.redblobgames.com/blog/2025-08... just for fun! :)

NeurIPS has decided to do what ICLR did: As a SAC I received the message 👇 This is wrong! If the review process cannot handle so many papers, the conference needs yo split instead of arbitrarily rejecting 400 papers.

Where do some of Reinforcement Learning's great thinkers stand today? Find out! Keynotes of the RL Conference are online: www.youtube.com/playlist?lis... Wanting vs liking, Agent factories, Theoretical limit of LLMs, Pluralist value, RL teachers, Knowledge flywheels (guess who talked about which!)

How astronauts control robots from space (featuring our quadruped Bert 👀) youtu.be/BMFPVCu16SQ

This one's been a long time coming. In this post on Decisions & Dragons I answer "Should we abandon RL?" The answer is obviously no, but people ask because they have a fundamental misunderstanding of what RL is. RL is a problem, not an approach. www.decisionsanddragons.com/posts/should...

📣Registration for EWRL is now open📣 Register now 👇 and join us in Tübingen for 3 days (17th-19th September) full of inspiring talks, posters and many social activities to push the boundaries of the RL community!

If machine learning is a game, it’s Calvinball. Bitter lessons from chess don’t apply.

Join us for the 𝗖𝗼𝗥𝗟 𝟮𝟬𝟮𝟱 𝘄𝗼𝗿𝗸𝘀𝗵𝗼𝗽 𝗼𝗻 𝗢𝗽𝗲𝗻-𝗦𝗼𝘂𝗿𝗰𝗲 𝗛𝗮𝗿𝗱𝘄𝗮𝗿𝗲 𝗶𝗻 𝘁𝗵𝗲 𝗘𝗿𝗮 𝗼𝗳 𝗥𝗼𝗯𝗼𝘁 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴! The day will bring together researchers and makers from both academia and the industry to discuss open-source robot design, integration with reinforcement learning […] [Original post on fosstodon.org]

Which ideas in computer science don't do better with faster computers? And what is a "human inductive bias?" (I never know what anyone means by "inductive bias." It is a term that defies definition.)

Pierre-Luc Bacon: Beyond the Gym Interface RL4RS Workshop @rl-conference.bsky.social

🚀 As we prepare for the next era of space exploration, coordinating intelligent robotic teams will be essential. Surface Avatar lays the groundwork for astronauts and robots to explore the Moon 🌙, Mars 🔴 and beyond 🌌, together 🤝 📖 Read more: blogs.esa.int/exploration/...

🤖 Meet the robots! • Interact: @esa.int's wheeled robotic arm 🦾 • Spot: @esa.int's nimble robot dog 🐕‍🦺 • Justin: @dlr-en.bsky.social's humanoid robot 🤖 • Bert: @dlr-en.bsky.social's small robot dog 🐾 Together, they’re pioneering seamless human-robot collaboration for space exploration 🚀

🤖 Imagine controlling 4 robots exploring a Mars-like landscape… from the International Space Station 💫 Last week, NASA astronaut Jonny Kim guided @esa.int and @dlr-en.bsky.social robots in the final Surface Avatar experiment, testing human-robot teamwork for future space missions 🌕🚀 🧵👇

𝗧𝗿𝗮𝗰𝗸𝗶𝗼 𝗶𝘀 𝗼𝘂𝘁! 🚀 If you've been looking for a lightweight, open-source package to use for tracking metrics, parameters, and hyperparameters, look no further! 👩‍💻 Trackio is local-first, free and open-source, and comes integrated with 🤗 Datasets and Spaces 📈 huggingface.co/blog/trackio

SB3 v2.7.0 is out, adding n-step returns for off-policy algorithms via the `n_steps` argument. ```python from stable_baselines3 import SAC # SAC with n-step returns model = SAC("MlpPolicy", "Pendulum-v1", n_steps=3, verbose=1) model.learn(10_000) ``` More info: github.com/DLR-RM/stabl...

Starting in ~20 minutes

Get ready to watch an astronaut from the ISS 🛰️ control our robots tomorrow at 13:20 (Munich Time), one of them being our favorite little happy hopper Bert =) Link: m.youtube.com/watch?v=m4Y5...

Here is an #Upkie wheeled biped getting up on its own: https://www.youtube.com/watch?v=ZoFQhNnQ62U A combination of reinforcement learning to find an initial trajectory, manual adjustments to transfer to the real robot, and model predictive control for balancing.

Need for Speed or: How I Learned to Stop Worrying About Sample Efficiency Part II of my blog series "Getting SAC to Work on a Massive Parallel Simulator" is out! I've included everything I tried that didn't work (and why Jax PPO was different from PyTorch PPO) araffin.github.io/post/tune-sa...

Robot Randomization is fun!

There are some milestones to celebrate for Stable-Baselines3 (SB3) 🎉! - 10k+ stars on @github.com - 9M+ downloads on @pypi.org - 3300+ citations for the JMLR paper - 1000 citations for the SB2 repository I would especially like to thank the maintainers and our contributors!

The Impact of On-Policy Parallelized Data Collection on Deep Reinforcement Learning Networks thrilled to share our #ICML2025 paper led by Walter Mayor & @johanobandoc.bsky.social , with Aaron Courville, where we explore how data collection affects agents in parallelized setups. 1/

proud to share a survey of state representation learning in RL that my student ayoub echchahed and i prepared, that was just published on @tmlrorg.bsky.social ! this was the bulk of ayoub's masters thesis and he put a lot of work and care into it! a few details in thread below... 1/

RL Is Hard (Episode #35753) Debugging why PPO SB3 (PyTorch) was working where PPO SBX (Jax) was failing. Just one line of code different... 🙈 (the initialization of the actor network output)

Connect with the #robotics community on Bluesky with our Robotics Starter Pack:

After hearing this a lot: "no debugger = no Zed"—we’re excited to share that Zed's Debugger has finally launched! 🎉 zed.dev/debugger

New in SB3 (master branch): to use n-step returns for off-policy algorithms, simply pass `n_steps=3` (it will instantiate a `NStepReplayBuffer` being the scene). Our implementation does not use any for loops, only NumPy vectorised operations at sampling time (so the data container stays the same).

I collected some folk knowledge for RL and stuck them in my lecture slides a couple weeks back: web.mit.edu/6.7920/www/l... See Appendix B... sorry, I know, appendix of a lecture slide deck is not the best for discovery. Suggestions very welcome.

Implementing RL algorithms correctly is hard. Try to implement an n-step replay buffer and you might encounter off-by-one errors, invalid data sampling, or an incorrect discount factor. github.com/younggyoseo/...

I'm flying to Michigan today to present our new paper "A Pontryagin Perspective on Reinforcement Learning" at L4DC, where it has been nominated for the Best Paper Award! We ask the question: is it possible to learn an open-loop controller via RL? 🧵

Got some recent research related to 6D Object Pose Estimation? Want to present it at #ICCV2025 in Hawaii? 🌴 Then submit and present at the 𝟏𝟎𝐭𝐡 𝐈𝐧𝐭𝐞𝐫𝐧𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐖𝐨𝐫𝐤𝐬𝐡𝐨𝐩 𝐨𝐧 𝐑𝐞𝐜𝐨𝐯𝐞𝐫𝐢𝐧𝐠 𝟔𝐃 𝐎𝐛𝐣𝐞𝐜𝐭 𝐏𝐨𝐬𝐞 (𝐑𝟔𝐃). Paper deadlines: Jun 30 (in-proceedings), Aug 29 (non-proceedings)

(2) We have launched a new tool: Paper Skygest Bookmarks! This is a list of posts about papers that you have liked. You need to like the post *containing the url*; it doesn’t matter whether you like the post while using Skygest.

Today, we’re announcing the preview release of ty, an extremely fast type checker and language server for Python, written in Rust. In early testing, it's 10x, 50x, even 100x faster than existing type checkers. (We've seen >600x speed-ups over Mypy in some real-world projects.)

It's been a minute, but I've added a new Q&A to Decisions and Dragons! This one answers "Why is it better to subtract a baseline in REINFORCE?" Preview image attached. See the link for the full text. 1/2 www.decisionsanddragons.com/posts/why_is...

Coming soon from the team behind ruff and uv ( @crmarsh.com ): ty, an extremely fast Python type checker and language server, written in Rust. github.com/astral-sh/ty

After 7 years of inactivity, today I updated the C++ implementation of my simple serial communication to the latest libserial version. Code: github.com/araffin/ardu... Blog post: medium.com/@araffin/sim...

New blog post: Automatic Hyperparameter Tuning in Practice. Part I was about the challenges and main components of hyperparameter tuning (samplers, …). Part II is about the practical application of this technique to RL with the @optuna.bsky.social and SB3 libraries. araffin.github.io/post/optuna/

I wrote something up for AI people who want to get into bluesky and either couldn't assemble an exciting feed or gave up doomscrolling when their Following feed switched to talking politics 24/7.

Love RLC and want to host RLC 2027 in your city? We're soliciting hosting! Please help us get the word out! forms.gle/XAoU6ZkcVHUt...

« When software is open-source, it means it is open-source – that the source is open – nothing more. […] It does not mean open to contributions; It does not mean support is offered; It does not mean you’re entitled to feature requests; It does not mean the developer owes you their time; […] »

Discussion on the misconceptions of open-source software and the entitlement some users feel towards developers and their contributions. By @vale.rocks Found via @erogol.com vale.rocks/posts/open-s...

A thread of thoughts on radiance fields, from my keynote at 3DV: Radiance fields have had 3 distinct generations. First was NeRF: just posenc and a tiny MLP. This was slow to train but worked really well, and it was unusually compressed --- The NeRF was smaller than the images.

Llama 4 is out, and I think it goes very much in the wrong direction. It's bigger, trained on more data, with more engineering tricks (per layer lr sweeps 🙄), exactly the opposite of what should be researched. We know big models work. The frontier is now to reach these performances with less. 1/2