Brandon Rohrer — Machine Learning in Production for Robots

Production Challenges

Brandon Rohrer shares the complexities of transitioning machine learning models from development to production, particularly in robotics. He explains that while the trained model might seem like the simplest part, the real challenge lies in ensuring all components work seamlessly together. This involves managing data requests, preventing service disruptions, and adapting to constant updates in software and hardware.

Machine learning, specific issues are, it's almost impossible to consider all the possible inputs you'll get.

--- Brandon Rohrer

Rohrer emphasizes the importance of periodically retraining models to adapt to changing real-world conditions, which is crucial for maintaining model accuracy and reliability 1.

   

Model Robustness

Ensuring robustness in robotic models is a significant challenge due to the unpredictable nature of real-world environments. Rohrer discusses the importance of breaking down problems into subproblems to manage complexity, though he acknowledges this approach isn't universally accepted. He highlights the difficulty of obtaining high-quality labeled data, which is essential for training effective models.

If you close your eyes to that, move ahead with poorly labeled data, then badness happens, and you get models that are worse than no model at all.

--- Brandon Rohrer

Rohrer also reflects on his journey from mechanical engineering to data science, noting how his background aids in understanding and solving these complex challenges 2 3.

   

Robotics Complexity

The inherent complexity of robotics makes it a particularly challenging field compared to other machine learning applications. Rohrer explains that unlike image or audio processing, robotics involves navigating novel environments and requires vast amounts of data to learn effectively. He notes that even simple tasks for humans can be complex for robots due to the need for precise data and the unpredictability of real-world scenarios.

The robot's never going to experience the same thing twice. You're never going to get exactly the same camera image two times in a row.

--- Brandon Rohrer

Rohrer emphasizes the potential of simulation to overcome some of these challenges, though it comes with its own set of pitfalls 4 5.