The last one hundred years has seen incredible advances in machines automating repetitive tasks with precision in a factory environment. Humans can now realistically multiply their productivity by many orders of magnitude with the help of computer-programmed machines. Unlocking the next few orders of magnitude in productivity improvements will require bringing machines out of the factory and into the real world, where human labor still dominates the economy, and where things are messy, disorganized, and unpredictable. Massive technological breakthroughs in hardware and software will be required to enable the next generation of flexible, robust, and intelligent automation.


Affecting change in the world is fundamentally limited by input (sensor) and output (actuator) interfaces. Humans often underestimate the fidelity and performance of sensors and actuators needed to solve problems. This is a natural bias because we view and manipulate the world with biological sensors and actuators that are orders of magnitude better than most things we can build. Unlocking machine intelligence is gated by the inputs the machine receives and the outputs it can produce.


Nearly all traditional business and data processing use cases can now be accomplished at low cost with commoditized compute hardware. However, the fundamentally real-time and embedded nature of machines that physically interface with the world demand highly efficient and specialized compute resources. Perception, behavior planning, and control algorithms require low latency to create intelligent and robust systems. Novel compute architectures will enable a revolution in performance for the intelligent machines of the future.


Data-driven machine learning algorithms have dramatically improved the state-of-the-art in computer vision and perception. Yet, challenges with validating black-box algorithms and uncertainty in their ultimate performance and robustness have slowed the adoption of end-to-end deep learning solutions. Successful systems engineering approaches, especially in safety critical applications, should leverage a mix of traditional algorithmic approaches with new techniques, and may still need a human-in-the-loop to handle fault scenarios or to provide guidance. Finding the right recipe for each application and the right level of automation while considering the vast menu of traditional and new algorithmic approaches will be required to create viable solutions to unsolved problems.

Are you a founder?

Let’s talk.

Contact us