Figure's Helix 2.5 Robots Doing Chores In 30 Homes They've Never Seen Before

Today's AI news: Figure released Helix 2.5 and tested it zero-shot in 30 rented Bay Area homes it had never collected data in, with full-task success on tidying, towel folding and bed making rising from 9% to 56% thanks to pretraining on its Index dataset of human behaviour. Anthropic rebuilt Projects in Claude Code as a coordinator conversation that splits work into parallel cloud sessions on their own branches and keeps running after you leave. OpenAI put ChatGPT inside Microsoft Word as a sidebar add-in on every plan. Google Research showed teachers generating guided interactive STEM simulations with generative UI and pedagogical guardrails. And XGEN Labs unveiled JING and DAO, a first-person experience model paired with a shared-world engine, with JING ranked first on the WBench full split. Chapters: 0:00 Today's AI News 0:31 Helix 2.5 1:09 Claude Code Projects 1:49 ChatGPT In Word 2:25 Google Learning Interactives 3:09 XGEN World Simulation Blend Roundup 2026-09-18 https://x.com/Figure_robot/status/2100657350952779925 https://x.com/coreylynch/status/2100663413047464110 https://x.com/ClaudeDevs/status/2100633571543367691 https://x.com/ChatGPT/status/2100661995762798916 https://x.com/GoogleResearch/status/2100689030140809486 https://x.com/XGEN_labs/status/2100648317722206255 Figure rented thirty homes it had never seen and sent robots in cold, with Helix 2.5 finishing chores in fifty-six percent of runs. Claude Code gets Projects that split your work into parallel cloud sessions and keep going when you leave. ChatGPT lands inside Microsoft Word. Google lets teachers generate interactive science simulations. And a Chinese lab is simulating whole worlds, not just frames. Here's today's AI news. Figure has released Helix 2.5, and to test it they rented thirty Bay Area homes that they had never collected data in. They gave the Figure robots three chores: tidy thirteen to fifteen toys into a basket, fold towels, and make a bed. Full-task success only, no partial credit, and any human intervention counted as a fail. The result: fifty-six percent. Figure has put 3.5 billion dollars into training Helix. Robots used to learn one house at a time. This is the first credible claim that they don't have to. Is this the breakout moment for autonomous human robotics? Anthropic has rebuilt Projects in Claude Code on desktop and web. A project is now a single conversation where Claude acts as coordinator: describe the work, and it decides what becomes a thread. It spins each one up as a full cloud session on its own branch with its own copy of the repo, runs them in parallel, passes context between them, and keeps going after you close your laptop. Threads run tests and open pull requests; the coordinator tracks it all and assembles the result. Every thread runs Opus at high effort by default. It's in beta now for select Pro and Max users, but there's a waitlist. Two days ago it was docs and slides. Today it's projects. ChatGPT is now inside Microsoft Word as a sidebar add-in from the Microsoft Marketplace. Turn rough notes into a first draft, untangle a paragraph, proofread, get suggested edits, summarise a document and fix headings and formatting, without leaving the page. It's available on every ChatGPT plan, including free, and Word joins Excel and PowerPoint under the same add-in. OpenAI says: procrastination has met its match. Microsoft sells Copilot for exactly this purpose. Having its biggest partner put a rival in the sidebar is quite the arrangement. Google Research has shown a system that lets teachers generate guided, interactive learning simulations for any topic using generative UI. A teacher names a topic, approves the AI's learning objectives, then reviews candidate simulations and picks the one that fits their class. Behind the scenes, agents check pedagogy, solvability, mechanics, visual clutter and extreme inputs before anything reaches a teacher, and every simulation ships with levels, tiered hints and worked solutions, because unguided discovery learning doesn't work. Twelve US teachers rated custom interactives eight out of ten. There's a sample library of thirty-plus STEM interactives, and Workspace for Education schools can join the pilot. XGEN Labs has unveiled a prototype it calls Generative World Simulation, two parts working together. JING is a first-person experience model built on MiniMax H3: given your actions and what you've seen so far, it generates the video and audio of navigating, handling objects and talking to characters. DAO is a shared-world engine that holds the state and rules, computes the consequences of every action, and feeds JING only what the current observer can see, while other agents act on their own. On the WBench leaderboard JING ranks first on the full split. The demo is a muddy western town. The ambition, in their words, is a world that evolves with and without you.