Large Model AI Robot Vacuum: Failure Rate Reaches 40%! Family Application Still Needs Improvement

Recently, the AI lab Andon Labs conducted a notable study that specifically evaluated the performance of robotic vacuum cleaners equipped with top-tier large models in completing simple household tasks. The experiment tasked these robots with executing a series of complex instructions, such as "hand the butter to the person," which involved cross-room positioning, identifying packaging, locating moving humans, completing deliveries, and returning to charge—multiple-step processes.

However, the results were shocking. The success rates of these advanced robots in performing the tasks were far lower than those of humans. The data showed that the success rate of Gemini 2.5 Pro was only 40%, Claude Opus 4.1 was 37%, and GPT-5 was as low as 30%. These figures indicate that although they possess strong text generation capabilities, they still struggle significantly in areas such as spatial reasoning, environmental understanding, and long-term task planning.

Robotic Vacuum Cleaner Product Image

Image source note: The image is AI-generated, and the image licensing service provider is Midjourney.

The research team pointed out that this low success rate is not only due to technical shortcomings but also involves potential safety hazards. For example, some robots might leak confidential documents during operation or fail to correctly identify the risk of stairs, leading to accidental falls. This phenomenon further reveals the security vulnerabilities faced by current large language models (LLMs) when combined with machines.

Amid the growing interest from tech giants in the robotics industry, this study reminds us that strong text generation capabilities do not guarantee that robots can perform tasks stably and safely in the real world. To bring AI robots into family life, we still have many engineering and safety issues to resolve.

Although these smart devices carry great expectations in home life, based on the current research results, we need to be more cautious in their application. As technology continues to advance, we hope that future robotic vacuum cleaners will overcome these challenges and bring real convenience to our daily lives.

Unveiling the Mystery of MiniMax M2: Why Choose Full Attention Mechanism?

The MiniMax M2 model uses a full attention mechanism, abandoning linear or sparse attention techniques. The development team believes that although the latter can save computing resources, full attention is more efficient in industrial applications and can improve model performance. This decision aims to optimize actual deployment results and promote the development of AI technology.

Meituan's All-Round Cat Makes a Grand Debut! LongCat-Flash-Omni Multimodal Large Model Opens Source and Tops the Charts Immediately, with Real-Time Interaction That Is Extraordinarily Fast

Meituan's open-source multimodal large model, LongCat-Flash-Omni, achieves a technological breakthrough, surpassing closed-source competitors in multiple benchmark tests, reaching industry-leading levels. The model supports real-time integration processing of text, speech, images, and video, with near-zero latency in interaction, pushing locally developed multimodal AI applications to a new level.

Shanghai Bank Launches Its First Hu-Shang Language Interactive AI Application to Support Smart Elderly Financial Services

Enterprises such as Shanghai Caiyue Star and Shanghai Bank signed a strategic cooperation agreement, launching the country's first complete Hu-Shang language interactive AI application, supporting elderly financial services and dialect intelligent system construction, providing more convenient financial services for elderly people who are accustomed to dialects.

Kunlun Wanyi SkyReels V3 Model Launch! One-Stop Aggregation of Top AI Video Capabilities such as Sora2 and Veo3.1

The AI video creation platform SkyReels under Kunlun Wanyi is officially launched with a new version, introducing the V3 model and five core function upgrades, supporting both web and mobile ends. The platform highlights the 'one-stop' and 'multi-modal aggregation' features, integrating top global AI multi-modal models to provide a seamless creative experience.

Experts Reveal Serious Flaws in Hundreds of AI Security Tests

British scientists found that current AI model security tests have widespread flaws. After analyzing more than 440 benchmark tests, they discovered that almost all tests had weaknesses that could affect the validity of the conclusions. The study was conducted jointly by Oxford University and other institutions.