One of the jobs robots are best suited for is the tedious and repetitive "pick and place" task common in warehouses, but humans are still much better. UC Berkeley researchers are accelerating the pace with a couple of machine learning models that work together to allow a robotic arm to plan its grasp and path in milliseconds.
People don't have to think hard about how to pick up an object and put it somewhere else - it's not just something we've been doing for years of practice every day, but our senses and brains are well-adapted for the task. Nobody thinks, "what if I took the cup, then pushed it very up and then to the side, then very slowly down onto the table" - the paths along which we could move an object are limited and usually quite efficient.
Robots, however, have no common sense or intuition. In the absence of an "obvious" solution, they have to evaluate thousands of potential paths to pick up an object and move it, which involves calculating the forces involved, potential collisions, whether it affects the type of grip that should be used, and so on.
Once the robot has decided what to do, it can be done quickly, but that decision takes time - several seconds at best and perhaps a lot more depending on the situation. Fortunately, the UC Berkeley robotists have come up with a solution that reduces the time it takes to do this by about 99%.
The system uses two machine learning models that work in relay. The first is a rapid-fire generator of potential paths that the robot arm has to take based on tons of sample movements. Create a set of options and a second ML model, trained to choose the best, chooses from them. This route tends to be a bit rough, however, and needs fine-tuning by a dedicated motion planner, but as the motion planner is given a "hot start" with the general shape of the route it needs to be taken, his final touch is just a moment of work.
If the movement planner worked on its own, it tended to take 10 to 40 seconds to finish. With a warm start, however, it rarely took more than one tenth of a second.
This is a counter calculation, however, and not what you would see in an actual stock situation. The robot in the real world also has to actually do the task, which can only be done so quickly. But even if the motion planning period in a real-world environment were only two or three seconds, reducing it to near zero is extremely fast.
"Every second counts. Current systems spend up to half of their cycle time on motion planning, so this method has the potential to dramatically speed up withdrawals per hour," said lab director and senior author Ken Goldberg. Perceiving the environment correctly also takes time, but is accelerated by improved computer vision skills, he added.
Right now pick and place robots are nowhere near the efficiency of humans, but small improvements will combine to make them competitive and, ultimately, more than competitive. The work done by humans is dangerous and tiring, yet millions of people do it around the world because there is no other way to meet the demand created by the growing online retail economy.
The team's research was published this week in the journal Science Robotics.