Introduction
Hello, everyone. This is the best shot. I'm Dafei. Google DMI has recently released a series of robot videos that showcase remarkable capabilities. These videos demonstrate a major transformation in robot technology, as DMI shifts the focus from hardware performance to a new robot plan. In this article, we will delve into the in-depth sharing of DMI's robot research manager Carolina Parada in DMI's broadcast to understand the technological revolution that is taking place.
Redefining Robots
In the understanding of most of us, robots are often associated with machine arms in factories or human-like robots in science fiction movies. However, the team led by Carolina Parada is redefining the essence of robots. They are not just making more flexible machine arms but building intelligent subjects that can understand, perceive, and act independently.
Technical Bottlenecks of Traditional Robots
This cognitive change has led to a deep investigation into the technical bottlenecks of traditional robots. Even the most advanced industrial robots can only perform fixed tasks in a predetermined environment. When faced with unfamiliar scenarios, they often struggle. In the 2010s, the mainstream method was to train robots through strengthening learning. For example, in DeepMind's early experiments, robots learned to balance by stacking objects. However, this method has a fatal flaw. When the shape or placement of the stack changes, the robot needs to learn again for months.
The Key Turning Point in 2022
In 2022, a key turning point occurred. The team introduced the large circle model into the robot system for the first time. When Carolina said to the robot, "I'm thirsty," the machine was able to automatically complete the one-touch action of detecting the drinking water machine, taking the cup, and taking the water. This marked a significant evolution of robots from sensor-driven executors to language-understanding thinkers.
Gemini Robot Technology
Now, DMI's Gemini robot technology essentially converts Google's multi-modal large model into the operational capability of the physical world. In the classic experiment of catching bananas, the robot uses only dual-lens visual cameras and the Gemini model to complete the operation. It first identifies the color, shape, and space of the banana through the visual language model VLM and then calls the object capture knowledge library for pre-training. This knowledge library is based on tens of millions of object capture data and provides comprehensive strategies for different types of objects.
The Guanlan Experiment
The guanlan experiment is even more shocking. When the creative team entered the lab with a mini blue frame and small ball that the machine had never seen before, the researchers did not conduct any specialized training on the robot. But with Jimny's robot, it took only 200 milliseconds to complete the entire process from visual recognition to action planning. The robot first understood that basketball is a ball that needs to be inserted into a circular frame through a multi-mode model. Then, based on three-dimensional space positioning, it calculated the angle and strength of the throw and finally completed the basket with a success rate of 92%.
Embodied Thinking
Carolina emphasized that this is not a simple model match but a real concept shift. Robots extract the abstract relationship from the head to the target from Internet-level documents and image data and apply it in a new scene. DMI's giant thinking, embodied thinking, aims to allow robots to gain the ability to understand the physical world of humans. For example, when organizing the desk, the robot needs to complete a complex cognitive chain, including border frame identification, language connection, and movement planning.
Robot Control System
Inspired by Daniel Kahneman's theory of System 1 and System 2, Dingmai designed a unique robot control system. The slow system of the cloud, corresponding to System 2, is responsible for running a complete Gemini multi-modal model and complex reasoning and long-term planning. The fast system of the body, corresponding to System 1, is responsible for deploying lightweight models, real-time processing of sensor data, and adjustment actions. This dual-system structure shows outstanding performance in dynamic environments.
Technological Innovations to Improve Data Efficiency
Traditional robots usually rely on a large number of wrong data during training. However, DayMind has improved the data efficiency by two levels through three major technological innovations. The first is the efficient use of human demonstration data. The second is the double-directional transfer between the real and the real. The third is the limited capacity of multi-mode pre-training.
Multi-Layered Security System
As the robot moves from the lab to the home, DingMind has built a multi-layered security system. On the physical level, DingMind robots use industrial-level force-controlled sensors. On the language level, the research team developed an Asimov database. At the system level, for extreme scenarios such as network interrupts, DMI has developed a system model.
Future Outlook
When asked when we will be able to have a robot assistant like Stargate, Carolina maintained an optimistic attitude. Although the robot is still like a 2- to 3-year-old baby and lacks continuous reasoning ability in complex scenarios, technology fusion is accelerating the breakthrough. In terms of social intelligence, continuous learning, and environmental modeling, significant progress has been made.
Conclusion
At the end of the broadcast, Carolina once again emphasized the essence of technological development. We are not making more advanced machines but expanding the boundaries of intelligence. When robots can understand why we need to organize the room, we have truly created the intelligent main body of the physical world. This is perhaps the most exciting future of robot technology.
Thank you for watching this video, and I'll see you next time.
{# 
{# 
{# 
{# 
{#