Can an LLM agent play Angry Birds? Introducing Embodied Playground 🤖 410 simulated task environments that test physical, strategic, and spatial reasoning. 🌏 Website: winniechen2002.github.io/Age… 📄 Paper: arxiv.org/abs/2610.27039 💻 Code: github.com/Winniechen2002/Em…

🚀T-WLA: Visual memory for fast touch 🖐️Memory mechanism connects vision and touch ⚡6× T-Rex’s decision rate (A6000 Pro) 📉30% lower finger-joint error 🤖2× T-ReX’s average success rate across 4 real-robot tasks 🎬 arxiv.org/pdf/2610.28814 @MIT @MIT_CSAIL @medialab @Harvard

This is a platform to evaluate how well current frontier foundation models perform on tasks that require spatial and physical intelligence. More models and tasks will continue to be added and evaluated here: winniechen2002.github.io/Age…

🚀SteerAct: Stanford × MIT x Harvard 🧠24 test-time rollouts/task. Only steering one feature. 🔒Frozen VLA/WAM. 🤖Real-robot mean success: 31.1% → 52.2% (+21.1 pp; ~68% relative). 🎬See it in action: arxiv.org/pdf/2610.25632 @MIT @medialab @Harvard @Stanford #Robot #AI

Round of applause for Embodied Playground pushing the frontier forward with 410 simulated tasks challenging AI agents across physics, strategy, and spatial reasoning. Great work by the team! 🚀

Can we make dexterous robots compliant without sacrificing precise control? Excited to introduce FOMO-Dex: Force-Modulated Dexterity! 🖐️ FOMO-Dex brings force-aware compliance to dexterous teleoperation, enabling robots to pick up thin cards and chips 🃏, cut paper 🔪, turn nuts 🔩, and use screwdrivers 🪛. These force-aware demonstrations also train better autonomous policies! 🎥 Videos & paper: yingyuan0414.github.io/fomo-… 🧵👇 (1/8)

A “Slow Startup” Pledge for Professors You ask students to stop chasing paper counts, resist peer pressure, and embrace “slow research.” We’d like to return the favor. We’ve noticed a worrying trend: more and more professors rushing to register companies, produce a demo, and pitch a top-tier fund. Their profiles list an impressive collection of Founder, Co-Founder, and Chief Scientist titles. But ask about course content, assignment design, or students’ learning difficulties, and the answer is remarkably consistent: “Please contact the TA.” We understand your anxiety. When everyone around you is fundraising, not fundraising can feel like falling behind. When the professor next door is disrupting a trillion-dollar market, explaining gradient descent to undergraduates can feel like a questionable use of your life. But please trust us: this fixation on valuation is damaging your most valuable skill—the ability to actually be a professor. You worry that undergraduates are joining research labs too early. But some of them aren’t leaving the classroom because their intellectual curiosity has outgrown the curriculum. They’re leaving because the classroom has been abandoned. The slides are years old. The assignments come from the TAs. The questions go to the PhD students. The professor is attending a closed-door summit on “the future of education.” Students who want to learn something current are told: “We can’t cover that in class. Join the lab.” So they join. Then the group meeting gets canceled. You’re fundraising. Students didn’t leave the classroom first. You did. Let us therefore make our evaluation criteria clear. Company valuation is not our primary measure of a professor’s quality. We want to know whether you understand what you teach, can explain the assumptions behind a formula, and can work out why students are confused without concluding that “this cohort has weak fundamentals.” One course you can actually teach is worth more than ten companies you’ve merely lent your name to. We’re not against professors starting companies, just as you’re not against students publishing papers. We simply ask for consistency: if students cannot substitute publication counts for understanding, professors should not substitute funding rounds for teaching. A Series B is not an answer to a question about slide 12. Launching a “world-leading general intelligence company” in three months does not necessarily demonstrate depth of thought. The direction may have come from investors. The code may have come from students. The demo may have come from a PhD student’s third consecutive all-nighter. The part you fully understand may be the cap table. When a student cannot explain their personal contribution to a paper, you question their capacity for independent research. When a professor cannot explain which parts of their company were not built by students, we would like to apply the same scrutiny to their capacity for independent entrepreneurship. We would also appreciate some consistency in your philosophy of industry experience. When a student interns at someone else’s startup, it is “a distraction from research.” When they work for yours, it is “exposure to real-world problems.” When they pursue their own interests, they “lack focus.” When they rebuild a demo overnight because an investor requested a different use case, they are “learning to tackle open-ended challenges.” Apparently, the educational value of industry depends on who owns the equity. The same project cannot be “research training” when you assign it, “proprietary company IP” when you pitch it, and “research training” again when the student asks to be paid. That is not a training program. It is a business model. Students should certainly learn to live with uncertainty. But that uncertainty should concern whether an experiment will work—not whether today’s work contributes to their dissertation or your next funding round. Education is not a staffing strategy. Its purpose is to help students develop the knowledge and judgment to eventually work without depending on you. That takes time. You have to update slides, explain a basic concept more than once, and give careful feedback on work that will never appear in a pitch deck. You have to tolerate long periods of “no commercial return.” Real teaching is sometimes profoundly unsexy. It does not announce a breakthrough every week. When students fail to understand linear algebra, you cannot simply pivot the course toward “the foundations of general intelligence.” And before declaring that you are reinventing education, you should probably teach the class you already have. If students must join your lab to learn current material, take on your projects to receive feedback, and join your startup to see their advisor, the lab has become a premium upgrade to the education they already paid for. Except the second tuition payment is collected in labor. Slow down. There is still a class on Tuesday. As students signing this pledge, we therefore commit to evaluating professors by the clarity of their explanations, the patience of their feedback, and their potential to show up—not the number of funding rounds they have completed. We will treat the rapid production of near-identical startups with appropriate skepticism. We will prioritize actually learning something over becoming part of your startup’s talent pipeline. And we would like to offer you some reassurance. Please do not assume that everyone else’s fundraising means you are falling behind. Prepare your lectures carefully. Answer questions patiently. Mentor your students sincerely. Tenure will surely follow. You can absolutely get rich without starting a company. One course you have truly mastered is worth more than ten term sheets. Of course, we do not sit on tenure committees. We do not set university salaries. We have no authority to guarantee your financial security. But none of that prevents us from sincerely advising you to be less anxious, less transactional, and less preoccupied with what the system actually rewards. Do not let your colleagues’ funding announcements distract you. Wealth is merely a by-product of teaching. Education is the real goal. A professor who truly loves teaching will never be left behind. Please stop asking about the incentive structure. Have you tried loving teaching more? We have no proposal to change the incentives. But we’ve fixed your mindset.

Do you really need to finish every rollout in RL for LLM? Not if you make your critic better—and trust it! We propose Actor-Critic with Action Chunking (AC2): use a learned critic to score chunks of tokens => only need partial rollouts ⇒ train faster than GRPO!

🚀 Honeycomb: constant-size scene memory for video world models 🐝 Six fixed-size planes with recurrent read/write. 🔁 Better revisit consistency 💾 Constant storage ⚡ ~13 ms updates 📃 arxiv.org/pdf/2609.37690 @Harvard @NUSingapore @MIT #AI #ComputerVision #VideoGen