10 Reasons why AI has developed faster than Robotics
In the last ten years, we have witnessed an explosion in the field of artificial intelligence. AI is everywhere around us, on our computers, our mobiles, our cars. However, a field closely related to AI, robotics, has not developed as much. This post aims to shed some light on some of the reasons why AI has developed much faster than robotics.
ai; robotics
It has been an amazing experience to see the development of both robotics and AI in the last ten years. Having been a researcher in robotics, and having switched to the area of AI, I have witnessed the momentum AI has experienced in comparison with robotics. These are ten points I think contributed to the difference.
1. Human-level Performance
Perhaps the most important reason why AI has developed so much is that the level of performance is human-like. In 2015, ResNet from Microsoft surpassed the human-level performance in image classification, Microsoft again achieved the same milestone in speech recognition in 2016. That same year, AlphaGo from DeepMind, beat the world champion of Go and in 2018, Google's BERT, surpassed the human-level performance in several NLP tasks.
There is a certain number of super useful applications, like OCR, object detection, speech recognition, etc. that can be used by multiple tech companies to solve common day-to-day problems.
In contrast, the robotics field hasn't been that successful. Projects that one or two decades ago were pioneers in the field, like HRP, ASIMO, Willow Garage were closed. The only company that has achieved a human-level performance is Boston Dynamics, but the science behind their robots is not known by the scientific community.
2. Deep Learning
In 2012, a CNN trained on GPUs achieved a significant improvement in the image classification benchmark ImageNet, starting the era of deep learning. Quickly, the research community started to apply deep learning to other areas of computer vision, NLP, speech recognition, reinforcement learning and others. Month after month, the state of the art was overpassed. As of today, we haven't stopped the rate of improvement.
3. Private Funding of Research
Large tech companies like Microsoft, Google, Facebook and others started to invest heavily in AI. In 2014, Google acquired DeepMind and in 2015 OpenAI was founded. Some well-known researchers in the field like LeCun or Hinton joined tech giants, along with many other researchers.
The amount of funding has skyrocketed over the last ten years. In 2019, AI startups have raised more than $24B worldwide. That same year Microsoft announced that the number of employees working on AI would be 5000. Apple currently holds the record of the number of startups acquired, 20, followed by Google that acquired 14 and Microsoft that acquired 10.
The amount of funding in the field of robotics is much lower. Perhaps the biggest investor is SoftBank, which acquired two of the top robotics companies, Aldebaran Robotics and Boston Dynamics.
4. Software vs Hardware
Most AI practitioners use virtual machines, so if you break your OS, you just need to boot up a new machine and in 5min you are back to work. In contrast, breaking a robot component can be difficult and expensive to fix.
For AI, the internet and mobile devices provide data in high quantities. Robots get data through their sensors, and the information needs to be processed in real-time, providing commands to actuators and making sure that the battery has enough power.
Another important issue is that the hardware part in robotics requires innovations in areas with even less funding like materials, mechanical engineering, electronics or electrical engineering.
5. Opensource
Opensource was key to the development of AI. OpenCV and Scikit-learn were the pioneer libraries that paved the way for deep learning libraries like TensorFlow, PyTorch or MXNet. These libraries facilitated the proliferation and reproducibility of AI research.
The only comparable effort in robotics was the development of ROS, a robotics operating system. While ROS helped to boost robotics research, most hardware components are still proprietary, which slowed the development of the field.
6. Python vs C++
In 2018, Python became the most used programming language, and probably AI has a lot to do on that. Python vs C++ is the fight of usability vs optimality. Python made AI easy to use for many people, accelerating its adoption. C++ is still king in robotics, even though there is certain adoption of Python, people still use C++ due to its speed and low memory consumption.
7. Cloud Computing
Cloud computing has been one of the key drivers of AI. Nowadays, it is possible to compute large amounts of data relatively cheaply and without taking care of hardware maintenance. Three of the largest tech companies, Microsoft, Google and Amazon, have made significant investments in the cloud business. Microsoft and Google, are doing strong investments in AI research and development. They know that making advances in AI will get more and more companies to incorporate AI workloads into their processes, which has to be computed in a cloud.
On the contrary, robots tend to use their local computer to operate, so in many cases, the objective is to optimize the computation so the robot is able to react quickly in real-time. The bottleneck for robotics development is not cloud computing per se, it is the lack of a sustainable business that creates a profitable virtuous cycle such as the AI case.
8. Popular Culture
Pop culture has played an important role in AI and robotics development. While some stories like Star Wars, Avengers or the books of Asimov picture a world where robots and AIs are our collaborators and share our living space, others, like Matrix or Terminator, envision a future where machines rebel against humanity and try to destroy it. These stories have attracted many people to AI and robotics.
9. Company's Adoption
On a 2020 survey made by Forbes to 1500 industry decision-makers, 40% have at least an AI project implemented and 90% are planning to implement an AI project in the short term, if not already. According to this report, the market size of AI was $39.9 billion in 2019 and it is forecasted to reach $733 billion in 2027. These figures speak very clearly on the current and future adoption of AI by companies.
Regarding robotics, the traditional industrial robotics market has been active for more 60 years, especially in the automotive industry, and reached $41 billion in 2020. These robots haven't gone through significant innovations. On the contrary, the market of mobile robots have started to expand, initially by vacuum cleaner robots, valued at $2.5 billion in 2019, and more recently, robots for warehouses and logistics, as well as UAVs.
10. Job Opportunities
More investment in AI, in addition to higher company adoption, drives a boost in job opportunities. According to Glassdoor, Data Scientist has been ranked a top job opportunity year after year from 2016 in both US and UK. In parallel, the number of masters in AI grew as well as free alternatives. Platforms for AI competitions like Kaggle attracted a large number of amateur and professional AI competitors, according to this source, in 2020 Kaggle surpassed 5 million registered users.
In robotics, the number of job opportunities is lower, for many of the reasons explained above.
Future Challenges
In the area of robotics, there is a lot of space for improvement. Boston Dynamics is pioneering the era of service robots and they are probably investing in making their robots more autonomous. Also, new competitors may appear in the next years.
In the area of AI, the current trend of research will continue: more data is needed, which requires larger computation power to train larger neural networks. This will generate more intelligent algorithms, which will provide higher value to companies and citizens.
This continuous growth in AI could potentially plateau at some point, however, we haven't reached that point yet. Another possibility of big change is a new theory that doesn't require massive amounts of data to learn. In this case, it will be harder for tech companies to find the business justification to fund private research, so the virtuous cycle could be broken. There are already successful efforts towards that direction like few-shot learning, or maybe we need to go back to symbolic AI.
I would like to thank Víctor González Pacheco for reviewing the article and for his useful suggestions.