Joe Watson

phd researcher in robotics & machine learning for control @DFKI @ias_tudarmstadt @TUDarmstadt previously @DeepMind intern, @CMRSurgical, @Cambridge_Eng

[Joe Watson @JoeMWatson]

Congrats to this year's pioneers, the selection was extremely challenging!

[RSS Pioneers @RSSPioneers]

We are excited to announce the 2026 cohort of RSS Pioneers! This year’s cohort brings together an outstanding group of early-career researchers whose work spans the breadth of robotics. A heartfelt thank you to all the organizers who made this year’s program possible.

[Joe Watson @JoeMWatson]

Please join us! Room 211 in the Convention Center at 12

[Théo Vincent @Theo_Vincent_]

🌍World models can play an important role towards building general agents,

but what should be their role in decision-making?🕹️

@JoeMWatson and I are organizing a Social @iclr_conf on this topic🎙️

🗓️Feel free to join the conversation on Friday 24th April, at noon!!

[Joe Watson @JoeMWatson]

Come chat to us!

[Daniel Palenicek @DPalenicek]

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

[Daniel Palenicek @DPalenicek]

🎉 Really excited, our paper "XQC: Well-conditioned Optimization Accelerates Deep Reinforcement Learning" has been accepted at #ICLR2026 .

If you are interested in reinforcement learning, sample-efficiency, compute-efficiency go check it out. See you in Rio!

[Daniel Palenicek @DPalenicek]

🚀 New preprint! Introducing XQC— a simple, well-conditioned actor-critic that achieves SOTA sample efficiency in #RL
✅ ~4.5× fewer parameters than SimbaV2
✅ Scales to vision-based RL
👉 https://t.co/3xgo4tpbAK

Thanks to Florian Vogt @JoeMWatson @Jan_R_Peters
@Hessian_AI @DFKI

[Kay Pompetzki @ICRA @kay_pompetzki]

Could geometric cues help improve goal inference in robotics?

We explore this question at #RLDM today | Spot 86.

Stop by if you're curious about bridging motion planning and intent prediction.

[Joe Watson @JoeMWatson]

Scaling sample-efficient RL often relies on artisanal architectures (extra LayerNorm, etc)

Daniel and Florian found a major issue with vanilla MLPs: the larger network weights slow optimization, so simply adding weight norm unlocks sample efficiency for much harder tasks! 🚀

[Daniel Palenicek @DPalenicek]

🚀 New preprint "Scaling Off-Policy Reinforcement Learning with Batch and Weight Normalization"🤖

We propose CrossQ+WN, a simple yet powerful off-policy RL for more sample-efficiency and scalability to higher update-to-data ratios. 🧵https://t.co/Z6QrMxZaPY

#RL @ias_tudarmstadt

[Joe Watson @JoeMWatson]

@pfau I see them all as being quite Ballardian, i.e., https://t.co/0XCVXraiqX

[Joe Watson @JoeMWatson]

@NatalyaLobanova https://t.co/iZbEsPpRpN

[Joe Watson @JoeMWatson]

@BarryPierce The US wasn't immune either
https://t.co/H7UGZmI8n4

[Joe Watson @JoeMWatson]

@sirbayes I think NVIDIA used "inference performance" to sell their GPUs and no one wanted to correct Jensen
https://t.co/kXuCi3UVjU

[Joe Watson @JoeMWatson]

This was a big collaboration, with the help of @theo_grune82772, @an_thai_le, @kay_hansel, @AHendawy19, @OlegArenz, @CarloDeramo , @MilesCranmer and @Jan_R_Peters, and Chen, Oliver, Will, Fabian, Tanmay and Martin who aren't on twitter

[Joe Watson @JoeMWatson]

We recently collaborated with @ABBgroupnews to survey the recent literature on physics-based machine learning ⚡️, with a focus on knowledge- and data-driven inductive biases 🔭

Check out the survey here 👇 and please reach out if we missed anything
https://t.co/FropEBJjKc

[Joe Watson @JoeMWatson]

Active learning! 🔥Entropy-regularized control! ⚡️Sequential Monte Carlo! 🎲 Make sure to check it out

[Hany Abdulsamad @HanyAbdulsamad]

#ICML2024
@sahel_iqbal is presenting our paper "Nesting Particle Filters for Experimental Design in Dynamical Systems" next week. A novel way to amortize sequential Bayesian experimental design:
https://t.co/EPWV5Zoz7q

Jointly with @AdrienCorenflos and @simosarkka
@FCAI_fi

1/*

[Joe Watson @JoeMWatson]

@breadli428 yes, but with TRPO. They use slightly different MuJoCo environments that have a horizon of 100-200 steps rather than 1k.

https://t.co/oJFtESxYwv
https://t.co/Qm2M6adnkD

[Joe Watson @JoeMWatson]

@rlfromlux @EugeneVinitsky Isn't this just an ablation study of the impact of approximations?

If a new algorithm works with NNs empirically, but doesn't work so well in an approximation-free setting, then the empirical results may be due to implementation details rather than the new algorithm

[Joe Watson @JoeMWatson]

@hbouammar It's true that covariate shift is harder to reason about when you have an internet-scale training set, but I think it's unsurprising that we've seen jailbreak prompts that are essentially out-of-distribution inputs, e.g.
https://t.co/gL7JFJZx2C

[Joe Watson @JoeMWatson]

I’m presenting CSIL at the poster session this morning, come find me at #1906! #NeurIPS2023

Sandy and I also open-sourced the implementation a few weeks back, you can find it at
https://t.co/EkMbmPAuwu

[Joe Watson @JoeMWatson]

Excited to share my internship project from my time at @DeepMind, looking at sample-efficient imitation learning using entropy-regularized reinforcement learning

TL;DR: do behavioral cloning (BC), get inverse reinforcement learning (IRL) for free! [1/6]

[Joe Watson @JoeMWatson]

CSIL has been accepted at @NeurIPSConf as a spotlight! ✨

Big thanks to my internship hosts Sandy and Nicolas at @GoogleDeepMind Robotics

We hope to share the code in the near future 🤖

[Joe Watson @JoeMWatson]

Excited to share my internship project from my time at @DeepMind, looking at sample-efficient imitation learning using entropy-regularized reinforcement learning

TL;DR: do behavioral cloning (BC), get inverse reinforcement learning (IRL) for free! [1/6]