UC Berkeley Daniel, one of the hottest future dire

  • Detail

One of the future directions? UC Berkeley Daniel has developed a low-cost robot that can be deeply "self-taught" based on AI

Blue robot, which is low-cost, safe, specially designed for AI and VR controlled by a robot research team at the University of California at Berkeley, undoubtedly gives many people a glimpse of one of the development directions of robot people in the future

according to data, global robot sales have doubled in the past five years, but today's robots deployed in factories and warehouses have almost the same performance and configuration as our robots decades ago. They are powerful and precise, but they are expensive to buy and dangerous for humans

(screenshot from: UC Berkely YouTube video)

blue looks a little like a child's immature robot drawing: it is made of bulky 3D printing parts, with a pair of humanoid robot arms with pliers, and each arm has 7 degrees of freedom

blue robot's own weight is only 8.7 kg, and its continuous payload is 2 kg. It can perform familiar daily activities in an unstructured environment, such as folding clothes and making coffee. It can support researchers to use VR for control. This process is very important for training AI robot algorithm. Let the operator wave his arm, and then wave his arm back and forth. While researchers manipulate robots through VR devices, AI algorithm can imitate and learn behaviors, which is equivalent to that human beings reduce the process of trial and error for robots and let robots quickly master a certain skill. It can also train the use of artificial intelligence to manipulate objects, which is still very rare in robots

Pieter abbeel, a roboticist in charge of the project and UC Berkeley Daniel, hopes to change this. He said that blue has been built from scratch to take advantage of its export share to emerging countries, which will also greatly improve the expertise and accumulation of recent AI improvements. The fact that AI is becoming more and more powerful gives us a chance to rethink how to design robots. Abbeel said that AI has made great progress in recent years, which makes robots more intelligent in software, but it has been standing still in hardware for many years. We need to develop new robot hardware for today's AI era. Reducing the cost is also very important to promote robot research

based on these ideas, more than 15 researchers from different fields in Berkeley robot learning laboratory have developed a low-cost quasi direct drive (Qdd) controller for blue robot, and constructed a complete design paradigm, which can achieve unlimited automatic control; Robots can support popular AI based control methods; They also made detailed consideration for the design of the robot itself and the production process of the robot. The volume of the 3D printer is getting smaller and smaller in order to reduce its cost

Abbeel explained that most robots currently used are powerful and accurate. Their actions are predefined, and they just repeat the same actions over and over again, whether it's screwing cargo pallets, welding carts, or fastening screws into smart corners

in contrast, robots in the future will be reactive and dynamic. They will be able to work safely with humans without interrupting or injuring them, rather than planning their actions in advance. They will use cameras and sensors to navigate the world in real time

if you look at traditional robots, their design revolves around the principle of very high precision and repetitive motion, abbeel said. But you don't necessarily need submillimeter repeatability. (be able to perform the same task again and again, and the movement difference is less than one millimeter.) Humans have no submillimeter repeatability. Instead, we use our eyes and touch to feel and complete the work through feedback

abbel and his team, researcher Stephen McKinley and graduate student David gealy hope blue can operate in the same way. It has a central vision module with a depth sensing camera, and its arm is controlled by a motor with a rubber band, which makes it flexible. If you push the industrial robot arm, it's like pushing a brick wall. But blue is more like a person in a crowded subway car: push it, and it will move to the side wisely

this enables blue to work more safely, but it is also suitable for research using reinforcement learning, which is an AI training method that has become popular in robotics. The working principle of reinforcement learning is to require agents to complete a task and give rewards when the task is completed. This is basically trial and error, on behalf of colleges and universities, enterprises and institutions: managers do not know how to achieve goals at the beginning, and then slowly self-study

Pieter abbeel believes that once robots master a skill through imitation learning, they can quickly evolve this skill through reinforcement learning, and then reach perfection, which is much more effective than ordinary programming or reinforcement learning

using traditional robots with reinforcement learning may be expensive. Their lack of flexibility makes them brittle and vulnerable. In addition, reinforcement learning takes time to produce results, and because robots are expensive, the rapid rise and accumulation of costs may soon make people retreat

in order to successfully carry out repetitive production tasks, traditional industrial robots usually have high precision and intensity, but this cannot guarantee the safety of human beings when working nearby, and also cannot be competent for more flexible work. Therefore, cooperative robots came into being in recent years, aiming to combine the repetitive performance of robots with human capabilities. At the same time, in order to work with human beings, cooperative robots are mostly designed to prevent pinch injury and collision. But the price of cooperative robots is often more than tens of thousands of dollars

this brings us to another area where blue may make a difference and expand its capabilities. Before the arrival of blue, Berkeley's research robot was PR2 built by will garage. It also had a pair of arms and pliers, but the production cost was expensive, about $400000. In contrast, the bill of materials price of blue is only $3000. Abbeel said that the team has not yet determined the final price, but they hope to target within the range of $5000. According to its official introduction, blue robot is a double arm robot with seven degrees of freedom, and the price is expected to be controlled below $2000 after mass production

this is possible when you are willing to give up sub millimeter accuracy, because you will realize that you do not need AI based control, abbeel said

many other research laboratories and start-ups are also targeting this new model, hoping to teach robots how to use artificial intelligence. Abbeel is the president of one of them, a startup called embodied intelligence. Kindred AI is a company that manufactures robots and can select items in the warehouse. Openai, a research laboratory founded by Elon Musk, has done similar work with robot hands, and Google is also exploring AI training for robots

however, some experts are skeptical about the attractiveness of blue to the industry and market. They noticed that it was no different from Baxter, another cooperative robot with arms and pliers. Last year, Baxter and rethink robotics, two robot star companies, went bankrupt, causing a sigh

ankur Handa, a robot researcher at NVIDIA, said that blue's pliers limited the range of tasks it could perform, and even with AI control, its accuracy would be problematic. In general, I don't think they offer anything particularly new. Handa added that blue robots are still a step towards making cheaper robots

however, since last year, the State Council has repeatedly held executive meetings specifically for railway construction, l optimistic about the future of blue. The robot is currently in small batch production, but abbeel hopes to expand its scale and eventually turn to Berkeley open arms for outsourcing manufacturing to achieve large-scale mass production. The first target customers will be research laboratories and universities, where robots are currently shared among teams, just like computers in the 1960s. Providing cheaper robots will make them available in a wider range of scenes, thereby increasing the output of robot research

it is reported that in 2017, Pieter abbeel, together with Peter Chen, rocky Duan and Tianhao Zhang, founded embodied intelligence (now company name:), aiming to develop AI software to help robots learn complex operations more easily and efficiently with the help of deep imitation learning and deep reinforcement learning

more importantly, abbeel hopes that blue can provide a blueprint for future home robots: low-cost, flexible and suitable for human use. This design is completely in line with our ideas, he said. There are still many challenges in the future, not like we think that this particular robot will enter every household. (but) this is a design paradigm that leads us in a new direction

Copyright © 2011 JIN SHI