Autonomous robots/cars – why aren’t we there yet?
Autonomous cars are the most well-known example of robot autonomy, with many big companies (Waymo, Nuro, Motional, or TESLA) still pushing autonomy limits. The capitalization of the autonomy market was valued at $76.13 billion in 2020 and is projected to reach $2,161.79 billion by 2030. But despite all the efforts and promises for years, autonomous robots operate only in restricted areas or with limited accountability. Let’s dive into what still blocks us and what we can do to make it a reality.
Highly specialized or general robots
In our daily life, we are surrounded by many specialized machines that either help us make a coffee or move us several floors up or down, but we are yet not seeing the rise of robots as predicted in many sci-fi novels. Most of the robots we currently utilize are highly specialized, making them dedicated to specific, usually repetitive tasks like pick-and-place on the production lines.
We would love to see robots that can freely move around the world, helping us in tedious but various operations that humans no longer want to do. Who would not want a household robot (like our Maddie Robot) to help with mundane tasks, including cleaning, doing the laundry, and being a helping hand overall in our day-to-day life? Autonomy is, therefore, the critical ingredient that could make our life easier and allow robots to become a part of our societies.
Robot autonomy is already among us
When considering autonomy challenges, we should look at similar applications where robot utilization is already happening on a much grander scale. The cleaning robots are examples of autonomous robots that made significant progress. A few years ago, these robots were getting stuck on almost any obstacle while returning to their base was rather wishful thinking than a reality. Nowadays, these robots are surprisingly robust, keeping the up-to-date map of the apartments, being able to relocalize, and finding a route around most of the obstacles. Even though our homes are static environments with few surprising changes, the robots are still getting stuck and have taken years to achieve this level of autonomy. In the end, if it fails, the robot gets stuck without causing significant damage to the property or home residents.
The path to autonomous cars
The scale of operation of autonomous cars and autonomous cleaning robots cannot be compared with cars’ operation being more complicated on several levels starting from the complexity of the vehicle itself and ending with the complexity of the world outside. The latter is one reason we still do not commonly see autonomous cars. Autonomous cars can deal with most typical cases, but there are always some new situations that also surprise human drivers. It is even hard to imagine what kind of rare cases might occur on the roads and even harder to answer if it is possible to create a dataset that can cover the full spectrum of these cases. Then these poorly handled cases are studied to solve the so-called long-tile of autonomy, which takes a lot of time.
Looking at the same cases, humans are still much better at dealing with these non-typical scenarios than robots while also not expected to be perfect. Human drivers cause many dangerous road situations or even collisions … but we no longer expect anything different while autonomous cars are expected to be perfect. This widespread conviction creates a discrepancy between what technology can offer and what was promised, which does not help bring the technology to the market.
Truth to be told, autonomous cars still experience more accidents on average than human drivers (9.1 driverless car crashes per million miles driven compared to 4.1 crashes per million miles driven). As it turns out, it matters how robots deal with different road cases as driving in a different style than humans makes robot less unpredictable and thus impact the overall safety on the roads resulting in most collisions being rear-ended collision. But making robots move more like humans makes them break some small rules like not wholly stopping at the STOP sign (as human drivers do all the time). Considering that driving culture differs between countries, maybe autonomous cars should also have AI to use honk to operate in India?
New research bringing new possibilities for autonomous driving
Without a doubt, the number of autonomous challenges exceeded overoptimistic expectations when we were at the beginning of the autonomous research. At the same time, we still observe new challenges that need to be solved.
So let’s dive deeper into what techniques are emerging that can help us scale our solutions, keeping in mind that most companies have an abundance of data. In contrast, they cannot scale to infinity with the number of autonomy engineers. The first tool that comes to mind is the simulation that allows verifying the full autonomous decision stack in prepared scenarios. But the reality gap between simulation and reality was (is?) a limiting factor in what parts of the autonomous stack we could verify. Quite recently, the computer vision community showed a great path with NeRFs that already looks like a game-changer in this domain. NeRFs can take a bunch of images and create a photorealistic 3D reconstruction that can be the basis of proper simulation. The autonomous community did not omit such innovation, with Waymo presenting a reconstruction of the San Francisco area from 2.8 million of images. NeRFs are just an example that shows that companies working on autonomy are already pushing the limits of what is known and can benefit from presented scientific developments.
This is just one of the examples when classical, mostly geometric-based approaches like SfM, in this case, are slowly getting substituted with trained methods. Some areas hold firm against the neural network revolutions, like visual-based SLAM solutions. But particular aspects of SLAM, like feature detection, description, and even matching, are already dominated by trained approaches like SuperGlue, which is already two years old but feels like forever.
While initial optimistic approaches to end-to-end networks failed, we see a new exciting trend with the classical structure of SLAM components being the basis for creating a neural counterpart. DROID-SLAM is a noble example showing that classical know-how can boost neural network components to achieve state-of-the-art results, beating even ORB-SLAM3 on the EuRoC MAV dataset. Neural network (or at least trained) components are the future as they reduce the need for manual parameter tuning that was time-consuming but critical to making first visual SLAMs work. Moreover, machine learning-based approaches scale well with data that companies already have gathered. Recently, we can hear more and more voices saying that we should have more trained components like planners or that our classical techniques are not yet on the level to solve the problem. One can even say that right now, everything new has to be differentiable to be able to train it.
Robot autonomy is the future!
Whether you are more inclined to classical or trained techniques, everyone agrees that the recorded data for autonomy is a crucial enabling factor even to consider making robots autonomous, and the investors value it. The race is still open, and the number of possible autonomous robot applications is primarily unlimited. Interestingly, data-driven techniques like neural networks should allow even smaller companies to efficiently compete with more giant corporations pushing the limits of what can be achieved. Let us know if you agree with our assessment of the current state of robot autonomy! We are eager to see who was right in the next 5-10 years, but one thing we are sure of – autonomous robots will be a part of our lives soon.