✨ I'm excited to share that our paper, 𝗙𝗢𝗖𝗔: 𝗙𝘂𝘁𝘂𝗿𝗲-𝗢𝗿𝗶𝗲𝗻𝘁𝗲𝗱 𝗖𝗼𝗻𝗱𝗶𝘁𝗶𝗼𝗻𝗶𝗻𝗴 𝗳𝗼𝗿 𝗗𝗮𝘁𝗮-𝗘𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁 𝗩𝗶𝘀𝗶𝗼𝗻-𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲-𝗔𝗰𝘁𝗶𝗼𝗻 𝗔𝗱𝗮𝗽𝘁𝗮𝘁𝗶𝗼𝗻, has been accepted to @ICML 2026 in 🇰🇷 next week. We have also released the code and will share the paper and repository links below. One of the biggest bottlenecks for scaling Vision-Language-Action (VLA) models is not model size - it is 𝗱𝗮𝘁𝗮 𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝗰𝘆. Unlike language or vision models, VLA models rely on data collected through real robot execution. E𝘃𝗲𝗻 𝗮 𝗳𝗲𝘄 𝗵𝘂𝗻𝗱𝗿𝗲𝗱 𝗱𝗲𝗺𝗼𝗻𝘀𝘁𝗿𝗮𝘁𝗶𝗼𝗻𝘀 𝗰𝗮𝗻 𝘁𝗮𝗸𝗲 𝘄𝗲𝗲𝗸𝘀 𝘁𝗼 𝗰𝗼𝗹𝗹𝗲𝗰𝘁, 𝗮𝗻𝗱 𝗻𝗲𝘄 𝘁𝗮𝘀𝗸𝘀 𝗼𝗿 𝗲𝗻𝘃𝗶𝗿𝗼𝗻𝗺𝗲𝗻𝘁𝘀 𝗼𝗳𝘁𝗲𝗻 𝗿𝗲𝗾𝘂𝗶𝗿𝗲 𝘀𝘁𝗮𝗿𝘁𝗶𝗻𝗴 𝗼𝘃𝗲𝗿. This makes data efficiency a central challenge for real-world Physical AI. This motivated a simple question: 🤔 𝗖𝗮𝗻 𝗮 𝗿𝗼𝗯𝗼𝘁 𝗮𝗰𝗵𝗶𝗲𝘃𝗲 𝗰𝗼𝗺𝗽𝗮𝗿𝗮𝗯𝗹𝗲 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 𝘂𝘀𝗶𝗻𝗴 𝗼𝗻𝗹𝘆 𝗮 𝗳𝗿𝗮𝗰𝘁𝗶𝗼𝗻 𝗼𝗳 𝘁𝗼𝗱𝗮𝘆'𝘀 𝗱𝗲𝗺𝗼𝗻𝘀𝘁𝗿𝗮𝘁𝗶𝗼𝗻 𝗱𝗮𝘁𝗮? One of our solutions is FOCA with three key components:  • 𝗙𝘂𝘁𝘂𝗿𝗲 𝗹𝗮𝘁𝗲𝗻𝘁 𝗽𝗿𝗲𝗱𝗶𝗰𝘁𝗶𝗼𝗻: learning latent representations of future actions to provide future-aware supervision during adaptation.  • 𝗔𝗰𝘁𝗶𝗼𝗻-𝗳𝗿𝗲𝗲 𝘄𝗼𝗿𝗹𝗱 𝗺𝗼𝗱𝗲𝗹 𝘀𝘂𝗽𝗲𝗿𝘃𝗶𝘀𝗶𝗼𝗻: exploiting imagined future observations as additional supervision, reducing reliance on expensive action-labeled robot demonstrations.  • 𝗘𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝘁 𝗳𝘂𝘁𝘂𝗿𝗲 𝗰𝗼𝗻𝗱𝗶𝘁𝗶𝗼𝗻𝗶𝗻𝗴: injecting a compact set of representative future tokens rather than all generated tokens into the VLA policy for scalable and effective future reasoning. We evaluated FOCA on LIBERO, RoboCasa, and 𝘮𝘶𝘭𝘵𝘪𝘱𝘭𝘦 𝘳𝘦𝘢𝘭 𝘳𝘰𝘣𝘰𝘵𝘪𝘤 𝘱𝘭𝘢𝘵𝘧𝘰𝘳𝘮𝘴, 𝘪𝘯𝘤𝘭𝘶𝘥𝘪𝘯𝘨 𝘴𝘪𝘯𝘨𝘭𝘦-𝘢𝘳𝘮, 𝘥𝘶𝘢𝘭-𝘢𝘳𝘮, 𝘢𝘯𝘥 𝘩𝘶𝘮𝘢𝘯𝘰𝘪𝘥 𝘳𝘰𝘣𝘰𝘵𝘴. Across diverse manipulation tasks, FOCA consistently improved adaptation under limited demonstrations. ⚡ In several challenging settings with only 𝗮𝗿𝗼𝘂𝗻𝗱 𝟭𝟬𝟬 𝗱𝗲𝗺𝗼𝗻𝘀𝘁𝗿𝗮𝘁𝗶𝗼𝗻𝘀, 𝘀𝘂𝗰𝗰𝗲𝘀𝘀 𝗿𝗮𝘁𝗲𝘀 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗱 𝗳𝗿𝗼𝗺 𝗮𝗽𝗽𝗿𝗼𝘅𝗶𝗺𝗮𝘁𝗲𝗹𝘆 𝟱𝟴% 𝘁𝗼 𝟴𝟰%. Compared with 𝘮𝘰𝘳𝘦 𝘵𝘩𝘢𝘯 𝘵𝘦𝘯 𝘳𝘦𝘤𝘦𝘯𝘵 𝘝𝘓𝘈 𝘮𝘦𝘵𝘩𝘰𝘥𝘴, including DreamVLA, ThinkAct, E0, and other recent baselines, FOCA achieved competitive or state-of-the-art performance, with particularly large gains in low-data regimes. Many thanks to all students and collaborators @an_thai_le @Mniepert 🔗 Project page: focavla.github.io/ 🐙 GitHub: lnkd.in/ezWQ5WJK #PhysicalAI, #ICML2026, #VisionLanguageAction #VideoWorldModel

✨ 𝐓𝐡𝐫𝐢𝐥𝐥𝐞𝐝 𝐭𝐨 𝐬𝐡𝐚𝐫𝐞 𝐭𝐡𝐚𝐭 𝐨𝐮𝐫 𝐩𝐚𝐩𝐞𝐫 𝐄𝐱𝐆𝐫𝐚-𝐌𝐞𝐝 [1] 𝐡𝐚𝐬 𝐛𝐞𝐞𝐧 𝐚𝐜𝐜𝐞𝐩𝐭𝐞𝐝 𝐭𝐨 NeurIPS 2025! (one of three other ones accepted this year 🚀 ) ✨ Over the past year, my collaborators and I have been exploring a fundamental limitation of 𝐦𝐮𝐥𝐭𝐢-𝐦𝐨𝐝𝐚𝐥 𝐥𝐚𝐫𝐠𝐞 𝐥𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐦𝐨𝐝𝐞𝐥𝐬 (MLLMs) on 𝐝𝐚𝐭𝐚-𝐡𝐮𝐧𝐠𝐫𝐲 tasks and how to overcome it. So how does it help with your work? 🔍 𝐈. 𝐊𝐞𝐲 𝐅𝐢𝐧𝐝𝐢𝐧𝐠: 𝐀𝐮𝐭𝐨-𝐫𝐞𝐠𝐫𝐞𝐬𝐬𝐢𝐯𝐞 𝐌𝐋𝐋𝐌𝐬 𝐀𝐫𝐞 𝐄𝐱𝐭𝐫𝐞𝐦𝐞𝐥𝐲 𝐃𝐚𝐭𝐚-𝐇𝐮𝐧𝐠𝐫𝐲 𝐔𝐧𝐝𝐞𝐫 𝐃𝐨𝐦𝐚𝐢𝐧 𝐒𝐡𝐢𝐟𝐭⁣ ⁣ We found that 𝐫𝐞𝐝𝐮𝐜𝐢𝐧𝐠 𝐩𝐫𝐞-𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐝𝐚𝐭𝐚 during domain adaptation 𝐭𝐫𝐢𝐠𝐠𝐞𝐫𝐬 𝐢𝐫𝐫𝐞𝐜𝐨𝐯𝐞𝐫𝐚𝐛𝐥𝐞 𝐟𝐚𝐢𝐥𝐮𝐫𝐞𝐬 𝐢𝐧 𝐚𝐮𝐭𝐨-𝐫𝐞𝐠𝐫𝐞𝐬𝐬𝐢𝐯𝐞 𝐌𝐋𝐋𝐌𝐬 - failures that 𝐟𝐢𝐧𝐞-𝐭𝐮𝐧𝐢𝐧𝐠 𝐜𝐚𝐧𝐧𝐨𝐭 𝐟𝐢𝐱. This reveals a critical weakness: current models learnt by purely auto-regressive models are not robust to domain shift.⁣ Joint work with great collaborators @cmuptx @james_y_zou , and my supervisor @Mniepert. Details below