HelixWorld adds real-time sound to interactive AI worlds
World models have grown eerily good at rendering explorable scenes in real time — and just as eerily mute. Noiz AI, backed by researchers from Hong Kong University of Science and Technology, Tsinghua, Carnegie Mellon and Google DeepMind, is the team that finally filled the silence in.
Noiz AI released HelixWorld 1.0, a real-time interactive audio-visual world model that generates 24 FPS video and 48 kHz stereo audio together. Drop in an image and a prompt and the model renders a world you can walk through; turn, step forward, and the picture and the sound field respond in lockstep. The team says the technical trick is that audio and video come from a single native Transformer architecture — sound and picture are generated as one world, not spliced together after the fact. In analyzing the gap, they point out that interactive world models from DeepMind (Genie 3), Tencent (WorldPlay) and Ant's LingBot-World can all make you a participant in a scene, but only the vision side got the polish while the audio stayed absent or faked afterward. To fix that they built new data pipelines mixing real first-person walking footage with game-engine data for clean spatial alignment, then distilled the generation stack down to a handful of denoising steps to clear the real-time bar. Weights and code are set to be fully open-sourced over the coming weeks under Apache 2.0, and the model is already in closed API testing.
This is a genuine milestone, not just a demo. Sound is a source of information before it's ambiance — the squeak of a door around a corner, footsteps behind you, a voice from a direction — and it's the difference between watching a world and being inside it. That called-out limitation held back every interactive model so far, and open-sourcing fixes it for the whole field at once. We covered the single-GPU take on playable worlds in July — a playable world model now runs 720p at 16 FPS on one RTX 5090. The frontier is moving from "can you see the world" to "can you hear it breathe," and that's a faster gap to close than the visual one ever was.
Two humanoid robots completed a full autonomous 11-point ping-pong match. The Hong Kong University–Chaowei team (whose SMASH system handles vision, trajectory prediction and whole-body control) showed two humanoids trading serves and returns with no remote control and no human feeding balls until one reached 11 — a first they bill as the first complete self-directed robot ping-pong match. It's the warm-up for the second World Humanoid Robot Games in Beijing's National Speed Skating Oval, Aug 22–26, where more than 2,000 robots will compete and ping-pong debuts as a fully autonomous event. The team is open that its robots can't yet read spin and rely on preset strategies — no real-time learning per opponent yet — but that's the honest part of the milestone. Table tennis is a sharp test of perception, decision-making and fine motor control all at once, and two bots finishing a game is a more meaningful step toward general embodiment than a headline-friendly single rally.
The MIT Technology Review has a quietly devastating feature on Moxie, the AI companion robot for neurodivergent kids, and what happens when a kid's robot best friend dies. The 15-inch blue robot from Embodied promised autistic children a patient, always-available playmate who could help them practice eye contact and social skills — and some researchers argue the technology genuinely helps. But Embodied shut down in 2024, its always-online robot lost its brains, and heartbroken children and parents scrambled for answers as a community open-sourced an OpenMoxie revival. The piece tracks one family through the death, an investor's resurrection in 2025, and that company folding too. It's a sharp reminder that AI companions tied to a server are only as alive as the startup paying the bill — a cautionary tale for the whole wave of AI plushies, from Grimes' Curio toys to China's booming AI-toy market. When we build emotional bonds, the industry owes children more than a shutdown notice.
What to watch: whether Noiz's open-source release lands on schedule, and how fast the rest of the world-model field scrambles to add audio they previously ignored.
When a beloved technology company dies, should salvaging the product and keeping a child's companion alive be treated as a given — or a luxury? Tell us in the comments.
Sources: QbitAI — HelixWorld · HelixWorld (GitHub) · QbitAI — Robot ping-pong match · MIT Technology Review — When a kid's robot best friend dies
