CyberRobo

Teach a robot a chore in 5 minutes You wipe your kitchen table five times while a camera watches. Five minutes later, a robot is wiping it for you. That's Steertuning, from Lerrel Pinto's group at NYU: 5 human demos, under 5 minutes, 71% success across 12 tasks in 12 new environments, including kitchens in real homes. This is a wheeled semi-humanoid, a YOR mobile base with two Wuji dexterous hands, learning straight from how people use theirs. Under it is a 400M-parameter diffusion policy pretrained on about 110 hours of first-person human video. Hands are erased from the footage and each scene becomes 5,000 3D points, so human motion transfers to robot hands. A new task takes 2 minutes of finetuning. At deployment, the robot picks the demo closest to the scene in front of it and uses that trajectory to steer its actions as they're generated. Gemini ER 1.6 plans the bigger steps. 71% is still a research number. But show it a few times, and it learns. That's how robots start giving people their time back.

Humanoid robots: I need more, much more data… I just watched an interview with the Memories AI team. One thing really stood out to me: by next year, Physical AI data collection could be entering the 10-million-hour range. The entire field is desperately hungry for real-world data. Just a few months ago, the egocentric data volumes being discussed across the industry were still around 1 million hours. More recently, we've been seeing figures to 2 million hours. But that's still not enough. Even 10 million hours may not be enough... Across the Pacific, at the World Internet Conference in Wuzhen, China, Li Shuai, deputy director of the AI Center at China's Ministry of Industry and Information Technology's Fifth Electronics Research Institute, put the gap into perspective: The core supply-demand gap in embodied AI data is roughly two orders of magnitude. Effective data stock is at the million-hour level, while leading companies estimate their demand at 100 million hours. It's pretty clear that demand for data is far ahead of supply at the upstream end of Physical AI. This is becoming a genuine must-have business. Look at what's happening-- On one side, companies are building dedicated humanoid robot data-collection and training facilities. China alone has more than 100 publicly disclosed embodied AI data centers and training bases. and like the NEURA's RoboGym with TUM in Germany, Apptronik's Robot Park in the US, and Tesla's Optimus Academy are all part of this push… On the other side, NVIDIA, Google DeepMind and World Labs are pushing world models, synthetic data and simulation to expand the training pool. NVIDIA generates synthetic robot trajectories from human demonstrations. Google's Gemini Robotics 2.5 combines real-world demonstrations with simulated practice, while World Labs' Marble generates 3D environments for robot training and testing. Then there's another layer: companies like Memories AI are building platform-based, one-stop data infrastructure to collect, organize and supply first-person real-world data at scale. These three routes may look like different technical bets, but I think they will increasingly converge.We're already seeing this in models like Google's Gemini Robotics 2.5, NVIDIA's GR00T 2 and Skild AI's S1. They all bring together different combinations of human egocentric video, wearable demonstrations, robot teleoperation and simulation. The boundaries between these data pipelines are becoming less distinct. And there's another side to this race(this sounds rather unhumanistic)--access to real industrial environments, frontline workers, skilled technicians and service professionals around the world. Their practical experience is an enormous source of knowledge that robots need to learn from. I've said this more than once… This really is the gold rush for Physical AI data. But data is not the goal.

Gemini Robotics 2.5 is now on the factory floor. Google DeepMind released the new VLA model. In a POC with Apptronik and Mercedes-Benz, Apollo 2 picked automotive brackets from bulk bins and sorted them into trays in the right sequence for assembly. It hit 96% success after a few weeks of post-training, with almost no drop from the lab to the line. The same model was also tested on 200+ tasks, including tying bows, serving food with tongs, and pipetting. Gemini 4 Argon handled most of the practice by turning real camera views into digital twins, running thousands of simulated runs, and reviewing failure replays to tighten the training.

1/ Twist a bottle cap with your eyes closed. Your fingertips already know when the cap is about to slip. Humanoid robots will need that kind of touch before their hands can do real work. Robbyant (@robbyant_brain) just released LingBot-Touch, a model that helps robot hands make sense of what they touch.

Bitter lesson on the robot side? Foundation models and dedicated robot models might be getting one from frontier models like Claude Fable 5.2, GPT-6 Astra, and Gemini Argon...(soon, Deepseek) Fable 5.2 hit 64.9% on RoboDojo’s bimanual sim benchmark -- more than double its 5.1 score of 27.6%, ahead of GPT-6 Astra at 29.8%, and past a lot of systems built for robot control. The arms solve a puzzle cube, unscrew a bottle cap, fit a gear, and draw on tasks the model hadn’t been explicitly trained for. Embodied work spent the last couple of years stacking specialized VLAs, WAMs, and manipulation data...General frontier models are now closing some of those gaps zero-shot with stronger reasoning and code agents.

I've shared quite a few examples of humanoid robots working in supermarkets, restocking shelves,picking products... But handling soft plastic bags, rigid cartons and slippery bottles requires more than visual recognition. Robots need to regulate grip force, detect slipping and adapt to changing contact conditions. Vision tells robots what they're picking up. Tactile feedback helps them understand how to handle it. This is the data gap @DaimonRobotics is addressing with TacClaw-Infinity, a large-scale Vision-Tactile-Language-Action (VTLA) dataset for dexterous manipulation. >Large Scale:10,000+ hours, 1,000+ task types and 1M+ manipulation trajectories. >High Diversity: 20+ research institutions and 50+ collection sites worldwide. >Omni-Modality: Integrating vision, tactile, language and action to capture both task context and physical interactions. Using handheld tactile grippers, Daimon collects human demonstrations without relying on specific robot platforms, while contact-centered annotation preserves the physical structure of interactions. Across three robot embodiments and two policy models, adding tactile data raised π0.5's average task success rate from 50% to 74%, with a maximum task-level gain of 44.7 percentage points. TacClaw-Infinity is part of Daimon's broader tactile stack, spanning sensors, data infrastructure and the Daimon-TWM world model. Tactile data may be less visible than vision, but it provides the physical evidence humanoid robots need to learn how the world actually behaves.

Amazon is developing a home robot unlike anything we've seen... According to CNBC, Amazon's new Personal Robotics Group is building around Fauna Robotics, maker of the Sprout humanoid, bringing together Fauna's team and much of the talent behind its original Astro project. The group also draws on Lab126, Amazon's consumer hardware R&D unit behind Kindle and Echo, along with robotics research talent from MIT. There's another side--> Its Frontier AI & Robotics (FAR) team is developing foundation models and full-stack robotic systems, with research spanning locomotion, dexterous manipulation, multimodal learning and sim-to-real transfer. Its current deployment focus is largely on warehouse operations. Together, these efforts bring different pieces of the robotics stack: Physical AI capabilities from FAR, humanoid hardware and engineering expertise from Fauna, and home navigation and human-robot interaction experience from Astro. Amazon is also hiring UX designers and field support engineers, suggesting a focus on turning these technologies into products people can actually live with. Amazon already operates more than one million robots across its fulfillment network. Bringing robots into homes presents a different challenge, with people, pets, furniture and unpredictable everyday tasks. So, what will the next generation of home robots look like--bipedal or wheeled? --- (video is Fauna's Sprout humanoid）