What’s Needed Next, Before The Robot Revolution

Today, the pervasiveness of AI is still fundamentally limited in robotics because every company is collecting their own fragmented (and often incomplete) data set — if any at all. Moreover, the large industrial robotic arm vendors do not easily provide users access to low level data such as currents, torques, etc., to properly train their industrial arms. Because of fragmentation and lack of raw data, it is nearly impossible to train robots to be intelligent and introduce new applications. When there is a common platform shared across all robotic manufacturers, then finally we can unleash the power of reinforcement learning and let our imaginations run wild. Simply said, there is no comparable to Apple with harmonious HW/SW integration, no Microsoft with an impressive ecosystem of hardware partners, and nothing as widespread and open-source as Android.

Once this is figured out I believe the next “Industrial Revolution” could be setting upon us sooner than we think. Automation and robot adoption could be the much needed boost our global economy needs.

We all know the comparisons of the robot revolution to the industrial revolution, the endless debates between immigration and technology, between cheap labor in China and expensive automation. We also know the flood of VC money and talent pouring into robotics, and all the hype — sometimes vaporware, and sometimes validated by giants like Amazon and Softbank.

Like the field of artificial intelligence (AI), robotics has been buzzed about for decades over numerous hype cycles, but has never fully delivered on its many promises. So why now? There are two under-appreciated technology trends: new open-source software and semiconductors running AI at the edge. Both are falling in cost, quickening the rise of robots, and will likely lead to a massive step function in adoption. We have seen this story has play out before in the field of machine learning— back-propagation in neural nets had been around for decades but it was the advent of the GPU that catalyzed increased the pace of innovation and coined the term “deep learning”.

Advances in computer chips, increasingly dense flash memory, and cross licensing of IP led to ARM which has brought us to present times. Now robotic systems do a very good job of mundane deterministic tasks but lack the flexibility to learn and perform higher impact probabilistic jobs in industrial settings.

Self-driving cars give us another clue. Once separate disciplines and industries — roboticists, AI/ML talent, and hardware ecosystem providers (semiconductors & sensors) came together to solve a single problem. We are now starting to see that first wave of that talent and technology migrate to other industries. This has been the catalyst that is allowing robots to learn faster, build other robots, and work towards democratizing the tech for a variety of industries and applications.

Robots are getting smarter faster

OpenRobotics, OpenAI, and Nvidia have been at the forefront of innovations in robotic software. OpenRobotics supports ROS and Gazebo, where ROS is a shared collection of frameworks and libraries for robotic software development and Gazebo is a simulator based on top of ROS. While ROS is far from perfect and has limitations as a production solution in safety-critical environments it is nonetheless a much-needed step in the right direction. OpenAI has been advancing the field of of reinforcement learning (among other areas), expected soon to have a huge impact on industrial robotics. Nvidia has built on top of ROS by creating the Astro AV stack that includes API’s for perception and localization. Moreover, they have introduced Isaac, a photorealistic robot simulator that may end up being one of the most powerful products released by the company. Advances in robotics (and self-driving cars) will be based on simulating a world that obeys the laws of physics but defies the laws of time (to allow quicker learnings).

Until recently we have mostly relied on the cloud to run deep learning in applications which has the suspected limitations of latency and requiring wireless connectivity. Novel new chip architectures have been developed to natively run neural nets in hardware. With these advances in semiconductors we can have natural language processing, computer vision, and various forms of AI right at the edge without any need to ping an Amazon or Google server. Nvidia has had a few headlines — introducing Jetson, which is their edge compute platform and open-sourcing their inference engine called the Deep-Learning-Accelerator (DLA). It is also not a coincidence this platform was released after Drive PX2 to leverage lessons, intellectual property, and talent from the self-driving world. Intel acquired Nervana and Altera, while Lux invested in Mythic to enable any device to have super intelligence at the edge.

Robots are building other robots

Tesla has been quoted that they are building the machine that builds the machine. This will likely be one of the next frontiers for opportunities in robotics. Collaborative robots are expected to increase 10x by 2020 while simultaneously help bring back manufacturing jobs by increasing workers productivity. Veo Robotics was started to help achieve just that and provide robots with intelligence and spatial awareness to allow them to coexist with humans in a workcell and factory. Further out in the future, increasingly digital factories will change the way we manufacture by maximizing factory floor space in the x, y, and z planes.

Robots for everyone

With the cost and time of experimentation falling, now both innovation and the ability to purchase these robots are no longer confined to large companies but are within the realms of startups. As an example, robotic bin picking at one point was considered the holy grail but now has nearly been turned into a commodity service with advances in computer vision systems. But that’s just one example, Embodied Intelligence — will bring a new application layer utilizing reinforcement learning outside of the labs of PhD students to the masses.

Looking beyond industrial robotics and going inside the home, iRobot has averaged sales of more than a million units of Roomba each year over the past three years. I expect more robots to be sold in the home; and as more autonomy enters vehicles then more robots will enter food preparation and delivery as seen by Zume Pizza. Embedded development platforms such as Raspberry Pi and Beaglebone created a maker community abstracting away development to foster quick tinkering with robots. Nvidia’s Jetson platform and supporting software environment allows for much of the hardware challenges to become mitigated and accelerates time-to-market, enabling shorter product feedback loops.

So what is missing? For years now the tagline has been “data is the new oil” albeit all the advances in this industry there is still a huge lack of data. Until there a ubiquitous horizontal platform to collect and share data, it will be hard to imagine a true robot revolution. ROS had promise of being that platform but has come up short as it has tried to make its way out of research labs and startup garages. Just as mobile phones became better at an exponential rate with the adoption of production ready operating systems such as Symbian, iOS, and Android, I would expect the same to happen to the robotics industry

Disclaimer: Companies Lux invested in are in bold.

Follow me on Twitter at @BrandonReeves08