🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

we have trained a humanoid transformer with large-scale reinforcement learning in simulation and deployed it to the real world zero-shot

Excited to share the recent results from Ilija Radosavovic, Tete Xiao, and Bike Zhang learning a humanoid locomotion policy in the simulation and zero-shot transferring to the real world! learning-humanoid-locomotion.github.io

New autoregressive humanoid transformer from BAIR: arxiv.org/abs/2402.19469 website: …anoid-next-token-prediction.github.io

(1/3) Bigger is better, but we couldn't handle it until now! Our lab unveils xT & S2, turning the tide on gigapixel image processing with smaller, scalable models. Efficiency meets precision, reshaping how we model high-frequency images...

(2/3) xT lets you model gigapixel images by introducing nested tokenization and the fusion of distinct vision and language models. xT gets you higher accuracy, less memory usage per pixel, and higher throughput while preserving local and global context: ai-climate.berkeley.edu/xt-website/

(3/3) S2 demonstrates that smaller models can closely approximate the features of larger vision models, challenging the necessity for scale in tackling complex images. Pre-training smaller, multi-scale models with S2 can beat large models on the same task: github.com/bfshi/scaling_…

New Agentic LLM model finds which frame to look at to handle complex video queries....

Berkeley alum or Berkeley affiliated and thinking of a startup? The House AI Accelerator just announced its partnership with Perplexity & its Founder/CEO. Apply by this Sunday at midnight. bit.ly/44pIcEj

See Amir Bar's excellent new work on learning from our furry friends.... Sometimes they know more than we do!

Our new VH benchmark shows existing models (including GPTo) fail on simple visual retrieval questions when placed in context with multiple distractors; in contrast, our retrieval augmented 7B MIRAGE model can find the signal from the noise and answer the question...

Super excited to announce our work on 'real-'world models! Our Conditional Diffusion Transformer (CDiT) has complexity linear in the number of context frames, can plan and navigate in known worlds, and can imagine outcomes in new settings!

What happens when vision🤝 robotics meet? Happy to share our new work on Pretraining Robotic Foundational Models!🔥 ARM4R is an Autoregressive Robotic Model that leverages low-level 4D Representations learned from human video data to yield a better robotic model. Berkeley AI Research😊

Super excited about our new work on pretrained 4-D robotic foundation models. LLMs learned with 4-D representations on egocentric datasets transfer well to real world tasks!

I'm very proud to be part of the Prompt AI team: check out our Seemour launch today!

Nexar’s Open Dataset & Crash Prediction Challenge Is Live on Kaggle! The competition runs until May 4, 2025, and we’ll share key findings at CVPR 2025. We’re excited to see what you build! Dataset: lnkd.in/e_2hkPdg Join the challenge: lnkd.in/eujrn_Ma