Interlocking AIs allow robots to pick out and place sooner than ever earlier than


One of the jobs robots are best suited for is the tedious, repetitive "pick and place" task that is common in warehouses – but people are still much better at it. Researchers at UC Berkeley are accelerating the pace with two machine learning models that work together to enable a robotic arm to plan its grip and path in milliseconds.

People don't have to think long and hard about how to pick up an object and put it somewhere else – it's not just something we've been practicing every day for years, but our senses and brain are well suited for the task. Nobody thinks, "What if I picked up the cup, then jerked it really far up, then sideways, and then really slowly on the table?" – The ways we can move an object are limited and usually quite efficient.

However, robots lack common sense or intuition. Since there is no "obvious" solution, you have to evaluate thousands of potential paths for picking up and moving an object. To do this, the forces involved, potential collisions, the effects on the type of handle to be used, etc. must be calculated.

Once the robot decides what to do it can be done quickly, but that decision takes time – a few seconds at best, and possibly a lot more depending on the situation. Fortunately, UC Berkeley robotics found a solution that cuts the time it takes to do this by about 99 percent.

The system uses two machine learning models that work in relays. The first is a rapid fire generator with potential paths the robotic arm can take based on tons of sample movements. A number of options are created and a second ML model, trained to pick the best, selects from among them. However, this path is usually a bit rough and needs to be refined by a special movement planner. However, since the motion planner gets a "warm start" with the general shape of the path that needs to be taken, this is the finishing touch and is just a moment of work.

Diagram showing the decision making process – the first agent creates potential paths and the second chooses the best one. A third system optimizes the selected path.

When the motion planner was working alone, it usually took between 10 and 40 seconds to finish. With a warm start, however, it rarely took longer than a tenth of a second.

However, this is a table calculation and not what you would see in an actual storage area. The robot in the real world must actually do the job, which can only be done so quickly. But even if the motion planning period was only two or three seconds in a real environment, the reduction to near zero adds up extremely quickly.

"Every second counts. Current systems spend up to half their cycle time planning motion, so this method can significantly speed up selection per hour," said lab director and lead author Ken Goldberg. Understanding the environment is also time-consuming, but it is improved with improvement Computer vision capabilities will speed up, he added.

Right now, pick and place robots are nowhere near as efficient as humans, but small improvements will make them competitive and ultimately more than competitive. People's jobs are dangerous and arduous, but millions are doing it worldwide because there is no other way to meet the demand created by the growing online retail economy.

The team's research will be published this week in Science Robotics magazine.


Katherine Clark