Equipped with 40 degrees of freedom, the A2 also features dextrous five-digit hands with built-in torque sensors for precise manipulation, while its face hosts an interactive display with integrated RGBD and fisheye cameras, as well as LiDAR sensors, enabling advanced perception in dynamic environments. A 700 watt-hour battery provides two hours of runtime, while its 60-centimeter turning radius ensures agility even in confined spaces. On top of this, by running newer AI models like deep-seek and integrating retrieval augmented generation, end users can finally build custom knowledge bases for tailored applications in business consultation, guided presentations, and more with relative ease.

And by achieving 96% accuracy in sound source localization, the A2 is built for noisy environments, combining facial recognition, lip reading, and voice command processing for seamless communication. Plus, the A2 features a proprietary action GPT system to generate human-like gestures and movements in response to voice commands, enhancing the realism and subtle creepiness of interactions while executing light operational tasks. Then there’s the robot’s mobility system that’s setting a new standard for humanoids. By using multimodal mapping, 3D SLAM, and vector flux planning algorithms, the robot achieves level 4 autonomous mobility, operating with low latency and minimal drift to maintain stability even in dynamic or cluttered environments.

And with peak torque of 512 Newton meters, the A2’s custom engineered integrated joints ensure reliable, lightweight performance while its reinforcement learning and a dynamic planning engine allow the robot to generate optimal trajectories in real time for maximum flexibility and agility. Together, these capabilities make the A2 ideal for environments requiring adaptive navigation and task execution. In terms of awareness, a suite of six high-definition cameras and 360-degree LIDAR provides comprehensive panoramic perception with no blind spots. The robot also supports the synchronized construction of dynamic and static semantic maps for intelligent obstacle avoidance and navigation.

On top of this, a three-layer safety monitoring system operates at the business, system, and hardware levels of the robot, including redundant backups between the computing and safety systems, as well as dual path control to the actuators. In fact, this system even meets performance level D safety standards to further ensure reliable operation in both public and private settings. And with its multi-platform control feature, the robot works with smartphones, computers, and remote systems, while its multifunctional standby station handles transportation, quick charging, battery replacement, and storage to reduce the reliance on human intervention.

As for the A2’s price, it’s expected to begin selling publicly this year for between $75,000 to $150,000, which would give the robot a daily cost of as low as around $205 per day, or $8.56 per hour. Plus, China’s DeepSeek just disrupted the entire AI industry, meaning that robots like the AGI Bot A2 may immediately begin leveraging its advanced reasoning abilities for real-time decision-making and complex problem solving. In fact, DeepSeek’s $671 billion parameter R1 model was reportedly trained for a fraction of the cost of similar models while offering industry-leading performance in chain of thought reasoning, coding, and mathematical tasks, making it a powerful tool for humanoid robots requiring high-level cognitive capabilities moving forward.

In terms of cost, DeepSeek’s R1 can run for $8 per million tokens, compared to GPT4’s $15 per million input tokens and $60 per million output tokens. But another Chinese company is also moving up the ranks, as the new startup Lumos Robotics just unveiled a video stress test of its upcoming LUS-1 humanoid. As for its known specifications, the robot is 170 centimeters tall and weighs 60 kilograms. Additionally, its body appears to be designed for consumer, hobbyist, and business applications with the LUS-1 featuring a robust outer frame that easily handles falls, pushes, and uneven terrains.

For now, the price is still unknown, but experts estimate that the cost of the robot will likely end up at between $40,000 to $70,000, while future peripheral additions such as dextrous hands will likely increase this figure by another $10,000 to $20,000. All the while, as more details become available, the humanoid robot market in China and the US specifically are predicted to heat up substantially over the course of this year. But there’s been another agentic AI breakthrough, as Anthropic and OpenAI were just defeated on over 10 graphical user interface agent benchmarks by China’s new User Interface Task Automation and Reasoning System, or UITARS.

The first key to its performance lies in its integrated perception, reasoning, grounding, and memory, which are fed into a single vision-language model for end-to-end task automation across different platforms, all without predefined rules or manual intervention. Another one of UITARS’ standout features is its ability to interact naturally with graphical user interfaces. Impressively, while most AI tools operate in restricted environments such as virtual machines or browser-based sandboxes, UITARS interacts directly with operating systems like Windows, macOS, Android, and web-based applications. For intelligence, UITARS relies on four core components, starting with its perception, which is used to process multimodal inputs, including text, images, and interface interactions, to create a comprehensive understanding of GUIs.

Additionally, UITARS continuously monitors dynamic interfaces and adapts to changes in real-time to ensure precise and efficient task execution. Its second component involves action, as UITARS features a unified action space that standardizes interactions across platforms, whether desktop, mobile, or web. The third component is reasoning, for which UITARS employs both System 1 reasoning, providing fast, intuitive responses, and System 2 reasoning, which enables deliberate, high-level planning. Additionally, the system can decompose tasks into smaller steps, reflect on its progress, and correct errors in real-time, ensuring reliable task execution. The fourth component is memory, where UITARS incorporates short-term memory to maintain situational awareness and long-term memory to retain historical interactions and knowledge.

Additionally, UITARS is designed for cross-platform interaction, supporting desktops, mobile devices, and web applications under a unified framework to execute multi-step tasks. As a result, UITARS can generalize effectively and handle diverse real-world scenarios. But what truly sets UITARS apart is its status as a native GUI model. Because unlike competitors including Google’s Project Mariner or OpenAI’s Operator, UITARS doesn’t require virtual environments or restricted sandboxes. Instead, it just interacts directly with operating systems to provide unparalleled control and analyzes raw inputs to dynamically adapt to changing tasks and interfaces. [tr:trw].