Tossing items is one of those natural actions we take for granted. That’s probably because we’ve been doing it since childhood, or rather since two million years ago.
It turns out it’s not as easy as it looks. According to robotic experiments done by Google, it requires a lot of practice.
The project’s centerpiece is Tossing Bot, a robot developed in collaboration with MIT, Princeton, and Columbia scientists. The robot arm has been teaching itself to pick and accurately toss random items.
How TossingBot works
Robots have been picking and moving objects for decades now. Most have been programmed to complete single, repetitive tasks. TossingBot is unique because it can differentiate between a variety of objects.
Like the human brain, it uses neural networks to identify objects and adapt to their various shapes. It might start with errors but improve accuracy as it performs consecutive tasks.
When you toss an object, your brain’s neural networks adjust for several variations. These include the object’s shape, size, density, and weight. The toss is then adjusted for all these variables.
By integrating physics with machine learning, TossingBot is able to not only pick and toss but improve intuitively. As opposed to past versions that had extremely mechanical movements, future robots will have more human-like gestures.
Through deep learning, robots can learn from experience instead of being programmed on a case-by-case basis. Although Google’s previous robots have been able to learn to grasp and push, tossing objects requires better knowledge of projectile physics.
TossingBot uses physics models as a foundation for its initial controls. It then utilizes AI principles to make necessary predictions and adjustments, which significantly increases its accuracy.
Neural networks allow the robot to adapt to more unpredictable real-world scenarios. This combination of physics and machine learning is known as residual physics.
According to Andy Zeng, a student researcher with the project, TossingBot has achieved 85% throwing accuracy using this hybrid model. It’s also able to grasp objects in the clutter with 87% reliability.
Implications and potential applications
Residual physics gives TossingBot the ability to adapt to situations it has never experienced before swiftly. After being trained with simple-shaped objects such as balls, wooden blocks, and markers, it performs impressively with various new objects.
With items such as fake fruits, it takes one to two hours to achieve the same level of efficiency as with the training objects. At over 500 objects per hour, TossingBot’s mean pick rate is already on par with humans.
Its biggest advantage is increased efficiency. Tossing objects is a faster way of moving them than picking and placing. It has massive potential for speeding up everyday processes.
One field in which this innovation can be applied is logistics. Warehouse operations involve a lot of picking and walking or driving over significant distances. A toss-and-catch system presents sizeable cost savings.
It could also be customized for disaster operations. An intelligent robot can sift through debris faster than humans. As an added advantage, it won’t be prone to human errors caused by fatigue, poor eyesight, and emotional attachment.
Because of its intuitive nature, a TossingBot inspired system will require minimal to zero human supervision. That’s because it won’t need to be programmed for each new task. It will use experience from past tasks to adapt to new expectations.
Although TossingBot shows great potential, it still has a lot of refining to go through. Google Robotics claims it can already toss objects more accurately than Andy Zeng’s team.
While this is impressive, it’s yet to be tested against professionals. These include athletes and other people whose occupations involve regular tossing and catching. Examples are package delivery guys and circus performers who have refined this skill over the years.
At the moment, TossingBot can only work with non-fragile objects. The team is yet to figure out how to safely toss items that can break easily. This is a big limitation because, in the real world, many objects require tender handling to avoid damage.
The robot also relies overwhelmingly on visual data to make decisions, which limits its potential. Integrating additional senses such as force-torque may further increase its reliability when handling new objects.
Given Google’s persistent interest in robotics, TossingBot is likely to benefit from the company’s wealth of experience in the field. Other than its tossing capabilities, the team will also likely work on a catching system which can safely handle fragile items.
Hopefully, future versions of this innovative robot will demonstrate more advanced problem-solving capabilities.
In the meantime, we can appreciate more the capability of the robots we already interact with, such as robotic vacuum cleaners and robotic lawnmowers.