Markus Wulfmeier
@m_wulfmeier
Large-Scale Robot Intelligence - Research @GoogleDeepMind European @ELLISforEurope - priors: @oxfordrobots @berkeley_ai @ETH @MIT
Imitation is the foundation of #LLM training. And it is a #ReinforcementLearning problem! Compared to supervised learning, RL -here inverse RL- better exploits sequential structure, online data and further extracts rewards. Beyond thrilled for our @GoogleDeepMind paper! A…
Gemini does! And so will Claude, ChatGPT, ... Maybe it's time for LLM tooling to provide feedback on reviews before submitting! share.google/Vp1i8GvVbMCGUy…
Anyone knows adam?
'Google processed nearly a quadrillion AI tokens across all its products last month, which is more than double the number it processed in May.' Quadrillion = 10^15 Googol = 10^100 Only x 10^85 missing. Maybe token is the right unit to measure here. theverge.com/command-line-n…
We're hiring @GoogleDeepMind APAC! 🇯🇵🇮🇳 Seeking experts in multilingual, multicultural & multimodal AI to help shape the future of our frontier models including Gemini. This is a unique opportunity to make impacts on billions of users.
@GoogleDeepMind India 🇮🇳 & Japan 🇯🇵 are looking for strong candidates in multilinguality, multicultural, & multimodality areas. RS Bangalore: job-boards.greenhouse.io/deepmind/jobs/… RS Tokyo: job-boards.greenhouse.io/deepmind/jobs/… RE Tokyo: job-boards.greenhouse.io/deepmind/jobs/…
Excited to share our new @icmlconf paper, "EvoControl" A hierarchical (bi-level) reinforcement learning (HRL) framework for long-horizon continuous control tasks. EvoControl. A slow (30Hz) high-level policy sets latent actions/goals for a fast (500Hz+) low-level policy that…
One more talk at the EXAIT workshop you should not miss: Wenhao Yu talks about foundations got scalable robot learning and #Gemini! @GoogleDeepMind share.google/mKZBKwoxX9RguK… Thanks to @sukhijabhavy for organizing!
Sadly unable to visit #ICML2025 but we have many fantastic colleagues and students joining! Want to learn more about agents, LLM fine-tuning, exploration, CoT, reach out to @thsschmied! Check out his @GoogleDeepMind internship paper at the EXAIT workshop!…
#Robotics is hard, and so is #AutonomousDriving! Massive congratulations to my friends at @Waymo for proving me wrong! There was a time a couple years back when I was getting skeptical about the scale of both technological and societal challenges for scaling autonomy. Other…
100 million real world, fully autonomous miles driven on public roads. That’s more than 200 trips to the Moon and back. Thank you riders.
Sadly unable to visit #ICML2025 but we have many fantastic colleagues @GoogleDeepMind and students joining! Want to learn more about exciting things happening in RL, inverse RL, and how to get robots to move, reach out to @MartinaZambelli this week!
Excited to be attending #ICML2025 this week — even more special that it’s happening right here in the city I now call home, Vancouver! Looking forward to connecting with friends and colleagues and catching up on all the great work being presented.
Sadly unable to visit #ICML2025 but we have many fantastic colleagues @GoogleDeepMind and students joining! Want to learn more about hierarchical RL, evolution, high frequency robot control, reach out to @samianholt this week! Check out his paper share.google/aJ7VytlcklpBOh…
Thanks for the shout-out, @m_wulfmeier! Thrilled to open-source the non-prehensile block reorientation setup from my @GoogleDeepMind internship, now part of MuJoCo Playground. By leveraging domain randomization and curriculum learning, we achieve smooth, compliant high-frequency…
Sadly unable to visit #ICML2025 but we have many fantastic colleagues and students joining! Want to learn more about agents, LLM fine-tuning, exploration, CoT, reach out to @thsschmied! Check out his @GoogleDeepMind internship paper at the EXAIT workshop!…
Google announced LLMs are Greedy Agents on Hugging Face Effects of RL Fine-tuning on Decision-Making Abilities
Independent on agreement of crossing the point just yet, we're getting close to cheap, large-scale evaluation. Scalable evaluation = scalable learning! What you can measure, you can optimize. Research as meta optimiser does this. RL, in the broadest sense, can automate it.
Reality is the ultimate eval Threshold were our robots learn more by themselves than from expert demonstrations (teleop) was just passed by in silence, but I suspect in hindsight it will huge
No reasonable LLM builds just on synthetic data, or just supervised data, or just RL .... I'd be very surprised if we'd converge to that type of limit in robotics.
Imagine there'd be no text online to train language models, how would you make LLMs possible?
We're looking for people to join us to work on Gemini Diffusion and help revolutionize language modeling! Details below: job-boards.greenhouse.io/deepmind/jobs/…
Excited to share what my team has been working on lately - Gemini diffusion! We bring diffusion to language modeling, yielding more power and blazing speeds! 🚀🚀🚀 Gemini diffusion is especially strong at coding. In this example the model generates at 2000 tokens/sec,…
Fully agree with "Grounding Intelligence in Movement"! This paper really brings home the importance of embodiment related modalities in AI - movement itself grounds intelligence for both machines & us. Segado et al @KordingLab arxiv.org/abs/2507.02771
Time to warm up to the perspective that robot actions are 'just' another modality for language models.
Large language and vision models alone don't solve the whole #robotics problem. But they surely have a massive impact on generalisation and robustness! New scenes, new backgrounds, new objects, new people, new language and audio.... Massive props to @ayzwah, @ColinearDevin,…
We took a robot to RSS in LA running our new Gemini Robotics On-Device VLA model. People interacted with the model with new objects and instructions in a brand new environment and the results were amazing!
Amazing to see the generality & dexterity of Gemini Robotics in a model small enough to run directly on a robot. Incredible speed & performance even in areas with low connectivity. Excited to continue this momentum to make robots more helpful & useful to people
We’re bringing powerful AI directly onto robots with Gemini Robotics On-Device. 🤖 It’s our first vision-language-action model to help make robots faster, highly efficient, and adaptable to new tasks and environments - without needing a constant internet connection. 🧵
Don't have a robot? Try our newest Gemini Robotics on-Device VLA in simulation! Or become a trusted tester and tune and adapt the model yourself! youtube.com/watch?v=nVMY3-…