Daily AI & Robotics Wrap: Humanoids Advance in Intelligence, Training, and Commercial Deployment
October 1, 2025 – The humanoid robotics sector is experiencing rapid advancements, marked by significant strides in artificial intelligence models, innovative training methodologies, and expanding real-world applications. Recent developments highlight a collective push towards more autonomous, adaptable, and cost-effective humanoid robots, positioning them for broader integration into industrial, service, and domestic environments.
Google DeepMind Unveils Advanced Vision-Language-Action Models and New Humanoid Entrants
Google DeepMind has introduced Gemini Robotics 1.5, its most advanced vision-language-action (VLA) model, alongside Gemini Robotics-ER 1.5, an embodied reasoning system. These models are engineered to propel robots towards general-purpose intelligence in physical environments. Gemini Robotics-ER 1.5 functions as a “high-level brain,” capable of planning multi-step tasks, comprehending spatial layouts, and even utilizing external tools like Google Search. It then relays step-by-step instructions to Gemini Robotics 1.5, which translates visual and linguistic inputs into motor actions while employing a “think before acting” mechanism for enhanced decision transparency.
In related news, California-based Kinisi Robotics has launched its KR1 humanoid robot, specifically designed to tackle repetitive and physically demanding tasks in warehouses and logistics hubs. The KR1 features dual arms for pick-and-place operations, a 15-kilogram lifting capacity, and an eight-hour battery life with hot-swappable batteries. Built on a wheeled base, it prioritizes smooth navigation on flat warehouse floors. The robot processes data locally, ensuring reliable performance even in low-connectivity settings, and uses imitation learning, allowing operators to demonstrate tasks for the robot to learn independently without coding. Additionally, the K2 humanoid, from an unnamed Chinese manufacturer, has been highlighted for its affordability at approximately $34,000 (RMB 248,000), a price point that significantly lowers the entry barrier for humanoid prototypes. The K2 boasts an 81.3% energy efficiency and an eight-hour battery life, featuring a straight-knee gait suitable for industrial environments.
NVIDIA Bolsters Robotics Development with Open-Source Physics Engine and Foundation Models
NVIDIA is accelerating robotics research and development through the release of new open models and simulation libraries, including the beta version of the Newton Physics Engine and the Isaac GR00T N1.6 robot foundation model. Co-developed with Google DeepMind and Disney Research, the open-source Newton Physics Engine is now available in NVIDIA Isaac Lab, providing researchers and developers with tools to create more capable and adaptable robots. Humanoid robots, with their intricate joints, balance, and movements, demand advanced physics engines for accurate simulation, which Newton aims to provide.
The new NVIDIA Isaac GR00T N1.6 open foundation model integrates NVIDIA Cosmos™ Reason, a vision-language model tailored for physical AI. This model functions as the robot’s “deep-thinking brain,” transforming ambiguous instructions into detailed, step-by-step plans by leveraging prior knowledge, common sense, and an understanding of physics. Isaac GR00T N1.6 also enables humanoids to perform simultaneous movement and object handling, offering greater freedom for tasks such as opening heavy doors. Leading robot manufacturers are evaluating Isaac GR00T N models for developing general-purpose robots.
Meta Eyes Android-Like Software Platform for Humanoid Robots
Meta is making a significant push into humanoid robotics, aiming to establish a foundational software platform akin to Android for smartphones. The company’s CTO, Andrew Bosworth, indicated that while Meta is developing its own “Metabot,” the overarching goal is to create a software blueprint that can be licensed to other hardware manufacturers. Bosworth emphasized that software development, rather than hardware, represents the primary bottleneck in advancing robotics. To address the complex challenges faced by humanoid robots, particularly in dexterous manipulation, Meta’s Superintelligence AI lab is collaborating on building a “world model” that facilitates software simulation for intricate actions like picking up a glass of water.
Seven-Eleven Japan to Deploy GenAI-Powered Humanoid Robots
Seven-Eleven Japan (SEJ) and Telexistence have forged a partnership to develop and implement generative AI-powered humanoid robots in stores, with deployment of the “Astra” model targeted for 2029. These robots will be equipped with a Vision-Language-Action (VLA) foundation model, designed to address rising labor costs and workforce shortages. The initiative aims for robots to handle routine in-store operations, allowing human employees to concentrate on tasks requiring human-specific skills. This collaboration integrates Telexistence’s extensive data collection platform, already utilized by its beverage restocking robot “Ghost,” with Seven-Eleven Japan’s network of over 20,000 stores.
China’s Humanoid Robots Embrace Vocational Training and Open-Source Ecosystems
China’s robotics landscape is witnessing a significant evolution, with humanoid robots transitioning from laboratory demonstrations to acquiring practical skills through “vocational schools” and open-source initiatives. Humanoid robots named Kuavo, for instance, are being trained using VR and motion capture systems at facilities in suburban Beijing to perform warehouse tasks such as returning empty crates, sorting materials, and packaging products. These robots have demonstrated high success rates, exceeding 95% in practical applications.
This rapid skill acquisition is supported by open-source ecosystems, mirroring China’s strategy in large AI models. Unitree’s AI engine enables humanoid robots to learn efficiently without repeated real-world trial and error, using single snapshots and planned motions to simulate and guide outcomes. Other key players like Shanghai-based Fourier have unveiled open-source humanoid robots, and Alibaba DAMO Academy has open-sourced its vision-language-action model and a robot context protocol. AgiBot also launched an open-source model, aiming to drastically reduce the data required for robot training, such as cutting the samples needed to teach a robot to pour water from over 10,000 to just 1,000.
Robots Learn from Themselves with USC Viterbi’s ReWiND System
A team at USC Viterbi has co-developed a novel robotic system named ReWiND that allows robots to learn complex tasks autonomously from a single video demonstration, adapting and refining their skills without direct human input. This breakthrough enables robots to self-teach tasks like folding towels or opening specific bins by continuously receiving “dense feedback” at every second of their movement, rather than just at the task’s completion. This approach significantly accelerates the learning process and enhances efficiency, laying groundwork for robots to perform household chores. The ReWiND system’s ability to learn from mistakes and adjust trajectories in real-time, even with variations in environment or language commands, represents a critical step towards more robust and adaptable robotic intelligence.