One human demonstration. Any multi-fingered hand. Zero-shot sim-to-real visuomotor policy. morphometricimitation.github… Collaborators: @he_siming @ckwolfeofficial @HaozhiQ @LeaMue27 Shankar Sastry, Claire Tomlin, @JitendraMalikCV

Formal verification × agents 🚀 Our IDS paper was accepted as a NeurIPS oral! We study how agents can co-evolve code + proofs, achieving 3x the success rate of Claude Code on distributed systems specs. IDS Paper: arxiv.org/abs/2605.23109 Code: github.com/skydiscover-ai/sk… And also check out our recent work on the next challenge: ensuring the spec itself captures human intent. 👇 nitter.net/istoica05/status/21009…

Tonight we kick off the 3rd season of the @UCBerkeley AI & Society Initiative series with an interdisciplinary panel on cybersecurity & AI focused on the recent Hugging Face OpenAI and related incidents, 5:30-7:00 PM in @BerkeleyCDSS Gateway Building. Co-organized this year by @irenetrampoline, @alsuhr, and me, the event is a co-production of Law + EECS/CDSS. Featuring - @cchio (Co-Founder & CEO of Coverbase.AI) - @hoofnagle (Faculty Director, Berkeley Center for Law & Technology / Professor at @BerkeleyLaw) - Starchy Grant (Principal Systems Administrator @EFF) Even if you can't make it in person, sign up to join the community and receive our substacks and the recording. - RSVP: lnkd.in/gcE6v7bC - Substack : lnkd.in/g6NY4Yqh - Website: lnkd.in/gc3xXA98 See below for a list of our fall lineup so far, with more to come! See you in Gateway or online!

Introducing TANGO 💃🏻, our #CoRL2026 work on whole-body VLA for humanoid navigation. TANGO maps RGB directly to 29-DoF actions, enabling coordinated whole-body navigation through cluttered 3D spaces. Trained entirely in simulation, zero-shot in the real world. 🤖🚀

Please help spread the word! We are recruiting multiple Postdoctoral Fellows as part of the recently launched Bakar Computational Biomedicine Initiative (BCBI) at UC Berkeley and UCSF. BCBI website: bcbi.berkeley.edu/ Apply by Nov 1, 2026: berkeley.infoready4.com/#fre… (1/n)

We just trained a 2.3B MoE Hybrid Mamba model that competes with a Llama-3.2-3B dense using less than 1% of the pre-training FLOPS. We may not have a lot of resources in academia but we make up for it with creative students who find ways to train innovative new architectures across whatever hardware they can find. If anyone wants to help, we would love to scale this research to the next rung of the training ladder.

We pretrained a 2.3B MoE (360M active) Hybrid Mamba-2 that lands within a few points of Llama-3.2-3B using <1% of its pretraining FLOPs. No dedicated cluster. The run hopped between H100s, A100s, V100s (yes, V100s) and TPU v5p/v6e on a single codebase. Meet Rigel 🧵

Interestingly, one attention head can be important across five ICL task families! Our original thread gave an in-depth mechanistic analysis of addition ICL (quick video recap 👇). We've now generalized this to four more ICL task families, spanning both arithmetic tasks and semantic tasks! Our paper “Understanding In-context Learning of Addition via Activation Subspaces” has been accepted to COLM 2026! 🎉 See more details in thread. (1/N)

How do we run RL with real-time chunking (RTC)? In this work we figured out how to use a small RL policy with a large robot foundation model, where the RL policy observed more recent images (due to faster inference) and steers the policy toward better behaviors! A fun collaboration with Siemens, led by Brian Zhu, Momen Khalil, Emanuele Poggi from Siemens and @ehharrison4 from Berkeley, with lots of amazing contributors!

Asynchronous VLA inference reduces inference delay, but breaks the Markovian assumption necessary for RL fine-tuning. How can we enable RL fine-tuning of VLAs with async inference? We introduce ARLI: Asynchronous RL with Intermediate Information! async-rl-intermediate-inform… (1/n)

Check out our paper at COLM 2026. We make (world-model based) controller language-instructable. It needs no manual labels with rollouts + our pos-hoc annotation method. The result is a flexible controller interface that any VLM/human can plug into. zinengtang.github.io/instruc…

ABC is fully out and headed to CoRL 2026! ABC-130K is the largest open teleop dataset to date: 3,500 hours, 130K+ episodes, 195 tasks, collected on an $8K bimanual setup. The full stack is open: hardware, training code, sim, and eval, so academic labs can do this research on equal footing. Beyond the data, we ran the baseline science for the community to stack on: • Sim-to-real correlation (r = 0.91 on task progress) • Which offline metrics matter? • Scaling law in compute (training steps and batchsize) • Conditioning the policy Project page: abc.bot/ Github: github.com/amazon-far/abc

Agent harnesses seem to make a lot of difference, at least for cost, even on open source coding benchmarks. Melissa’s research project digs into why.

Does a model need its native harness? We evaluated seven models across the Claude Code, Codex, and Pi harnesses and found a surprising result: harness choice had little effect on task success but a substantial effect on cost. Sometimes, a simple harness is all you need!

🚨 Call for participation: the Child–Robot Dexterity Challenge at the CoRL 2026 𝗦𝗣𝗜𝗡 𝗪𝗼𝗿𝗸𝘀𝗵𝗼𝗽! Bring your robot, or run on ours. Apply by Oct 5. 🧒🤖 Robots and a child will take on the same everyday tasks, side by side, live on stage. The demonstration offers a simple way to see what today’s robots can already do and where there’s still room to improve. We make it easy to join: - Bimanual YAM setup (MolmoAct 2 style) + a dexterous Sharpa hand setup, provided on stage. - Bring your own hardware (subject to approval). 🌐 Website: spin-workshop.github.io/#con… 🗣️ Speakers & panel from Stanford, Columbia, TU Darmstadt, and AMI Labs. 📍 Nov 12 · Austin, TX #CoRL2026

Does your Claude model really need Claude Code…? 🤔 We evaluate 7 models on Claude Code, Codex, and Pi. Three surprising findings emerge: 1️⃣Harness choice has little effect on task success rate, but can significantly affect the cost 2️⃣A simple harness can be competitive 3️⃣The native harness isn’t always the best. Millions of people are using coding agents, but the impact of harness choice remains unclear. (1/n) More details in the thread. 🧵

New paper evaluating LLMs' professional knowledge across occupations! We develop a scalable approach to benchmarking occupation-specific knowledge by turning online authoritative sources into test questions. Really enjoyed working on this with @abhishekn and @ShreyasK1609!

New work in @ScienceTM by Meghana Kamineni and a great team led by @ZhiYu_ACGT and @pnatarajanmd! We used deep learning-extracted features from imaging and genomics to understand the role of spleen in coronary artery disease.