SonyAI

From research project to the cover of Nature to a feature played by millions: Gran Turismo Sophy turns five. In five years, GT Sophy went from a hypothesis about deep reinforcement learning in racing to a living, evolving product inside Gran Turismo 7. Along the way, the team solved problems that didn't have textbook answers: teaching an AI agent sportsmanship, scaling from 4 cars to 500+, and shipping a research prototype as a real game feature in under a year. The timeline, the team and what comes next: https://bit.ly/4vHyxpF #SonyAI #GTSophy #GranTurismo #DeepRL #AIResearch

Before we zoom into summer, we wanted to share the June highlights you may have missed. 🏓Our autonomous robot Ace beat seven ranked professional table tennis players under official rules, including World No. 26 Miyuu Kihara and Olympic silver medalist Miu Hirano. The run is drawing coverage in the robotics press. 📃We brought six papers to CVPR 2026 in Denver; and Peter Stone keynoted on foundation models for embodied AI. ⚖️Alice Xiang made the case for data rights in TIME. We also took part in FAccT 2026 in Montréal. 🔗 https://bit.ly/4wh2aOd #SonyAI #AIresearch #Robotics #PhysicalAI #CVPR2026 #AIethics

We're honored to share that the Fair Human-Centric Image Benchmark (FHIBE) has been recognized as an Honorable Mention in the General Excellence category of Fast Company's 2026 World Changing Ideas Awards. 🌎 FHIBE is the first publicly available, consensually collected, globally diverse dataset for evaluating fairness in human-centric computer vision — built to help the AI community test for bias more rigorously and responsibly. This recognition reflects our belief that responsible AI starts with better tools for accountability. Thank you to Fast Company for spotlighting the bold ideas working to build a more equitable future, and congratulations to our fellow honorees. Read more about the 2026 World Changing Ideas Awards: fastcompany.com ⚖️To learn more about FHIBE, visit https://lnkd.in/g59tjgJj #FCWorldChangingIdeas #ResponsibleAI #AIEthics #FHIBE

Following the publication of our results in Nature, the Sony AI team behind Project Ace kept pushing. The goal: Defeat professionally ranked table tennis players with an autonomous robot. Between February and April 2026, Ace won matches against seven professional players with current or recent world rankings under official competition rules, including a victory over Miyuu Kihara, currently World No. 26 in women's singles and two-time Olympic silver medalist Miu Hirano. How did we get here? 🏓 Improvements spanned the full stack, from Bayesian Optimization for more competitive serves, to a scaled-up Deep RL policy network (4.5M weights) that consolidates nine distinct skills into a single model. ⚒️ Hardware upgrades reduced robot weight by over 4 kg and increased motor power, enabling faster movement and more aggressive spin. 👀 Perception latency dropped from 10ms to 8.5ms, and updated physics models improved sim-to-real transfer. Hirano, who lost to Ace, put it simply: "I think this robot has the potential to rank in the top ten among women." She added: "Please let me try again. It's frustrating. It's really strong." Read the full breakdown of results and technical improvements: https://bit.ly/440oyPM

Sony AI and collaborators are presenting ten papers at #CVPR2026 in Denver, June 3–7. The work spans generative modeling, face aging, video-to-audio synthesis, 3D scene understanding, domain-adaptive and evolving-world perception, visual token efficiency, and knowledge transfer—each targeting a different constraint on building AI systems that operate reliably at real scale. 👉 Read the full roundup: https://bit.ly/4obvwuD #SonyAI #AIResearch #CVPR2026 #ComputerVision #MachineLearning #GenerativeAI

Researchers from the Sony AI's Ethics team will be in Dublin for the IAPP AI Governance Global Europe conference to introduce FHIBE, the first human-centric AI evaluation dataset made with ethics in mind: https://bit.ly/3S8zYhC In this session, you'll: ✅ Understand the urgency around ethical & regulatory frameworks in AI ✅ Discover a practical tool to identify & mitigate model bias ✅ Get insights on curating datasets aligned with today's requirements FHIBE serves as an example of responsible protocols across the data lifecycle, from consent-based sourcing to annotation, privacy preservation, and bias evaluation. Don't miss this opportunity to learn from our leading AI researchers and developers. Event Details: Date: June 3 Time: 14:45 GMT Location: Dublin Register now: iapp.org/AIGG26 Sony AI Ethics Team Speakers: Austin Hoag, Staff AI Engineer Victoria Matthews, Senior AI Policy Specialist Wiebke Hutiri, Senior Research Scientist Lilly Taranto, Director & Senior Privacy Manager

What if you could re-shoot the camera path after you've already recorded? Vid-CamEdit makes post-capture camera trajectory editing possible, keeping scenes spatially coherent, even across large viewpoint changes. 👉 Explore the research: https://bit.ly/4tTmPpX 👉 Read the full blog post: https://bit.ly/3PpkNju

Are you a PhD working on gen AI evaluation, guardrailing or responsible AI? Join me and my amazing team for a 3 month internship and the opportunity to make a tangible impact on ethical AI creation & protection in the creative industries. Background in VLMs or NLP is a strong advantage. We are open to candidates in the EU, UK and US. Here’s the link to the EU application: https://lnkd.in/dKhmK-RX #aiethics #ai #evaluation #guardrails #protectingcreators Alice Xiang, Morgan Klaus Scheuerman, Shruti Nagpal, Yunhao Ba, Julienne LaChance, Sabine Wildemann, Justinas Mišeikis

ASR hallucinations are risky precisely because they look confident. This research reduces Whisper's hallucinations at the source, improving encoder robustness under noise and aligning decoder behavior to a clean-audio teacher. 👉 Explore the research: https://lnkd.in/gxi-6NTe  👉 Read the Blog: https://lnkd.in/gcCi6XDq

🔊 Sony AI has released Woosh, a sound effect foundation model built for the professionals who create the sonic worlds behind games, film, and interactive media. We sat down with some of the team behind it to find out how it was built and why it matters. Most generative audio models are trained on general public data.  Woosh was built differently. The model was optimized specifically for sound effects, trained on professionally curated libraries, and evaluated against the vocabulary and annotation standards that sound designers actually use. The result is a model that understands the difference between a sound effect and an audio scene, and performs accordingly. The public release includes text-to-audio and video-to-audio generation models, open weights, and inference code for non-commercial use. A plugin for digital audio workstations is in development, with support for variation generation, inpainting, and personalization planned as the ecosystem grows. 👉 Read the full interview with Hakim MISSOUM and Marc F. on the Sony AI blog: https://bit.ly/4wzj4Zk To explore the model weights and demo samples directly, visit: https://lnkd.in/gF5_QxGG And to access the Woosh-Flow Private, please visit: https://lnkd.in/gW2AwpvK #SonyAI #AIResearch #SoundDesign #GenerativeAI #AudioAI