# From Robotics to Recommender Systems

Miguel Fierro, Microsoft | MLOps Podcast | Episode 240 | 55:54
Hosted by Demetrios Brinkmann

Source: https://www.youtube.com/watch?v=wEmO00ADppw
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/from-robotics-to-recommender-systems
Published: 2024-06-11
Tags: classical-ml, multimodal, recommender-systems

## TL;DR
- Robotics often gets better results from physics-based control than from machine learning when the robot's motion can be described with exact equations.
- Sports teams can use computer vision to track players and the ball, while recommendation systems can support strategy, health monitoring, and performance decisions.
- Recommendation systems usually follow batch, real-time, or hybrid architectures, and their business value must be shown through experiments and financial results.

## Summary
Miguel Fierro compares machine learning with the physics-based methods used in robotics. Boston Dynamics, he says, built much of its earlier work around equations of motion and model predictive control, especially for stability and posture. A neural network can approximate a known inverse-kinematics formula, but that adds little when the formula is already available. He then describes sports analytics, where computer vision tracks players and the ball and can help teams study tactics, performance, and athlete health. The main part of the conversation covers recommendation systems. Miguel describes batch systems that precompute recommendations, real-time systems that score current user activity, and hybrid systems that retrieve candidates before reranking them. He is direct about the cost of testing: teams may need weeks for an experiment, and many experiments fail. His advice for winning leadership support is to bring measured results and express the return in dollars. Large language models can add recommendation algorithms or power conversational, personalized search interfaces.

## Key ideas
### Robots can rely on physics when their motion is already understood
[02:21](https://www.youtube.com/watch?v=wEmO00ADppw&t=141s)
Miguel says many roboticists do not need machine learning for core control problems. Boston Dynamics spent decades using physics, including Newtonian mechanics and model predictive control. That controller predicts the robot's movement in real time, reads posture from sensors, and stabilizes the system. Miguel describes stability and posture as one of robotics' hardest problems, and his own PhD work focused on posture control. Machine learning becomes more relevant for separate problems such as computer vision, which many robotics research groups handle independently from force, control, and stability.

### Actuators change what a robot can do
[07:40](https://www.youtube.com/watch?v=wEmO00ADppw&t=460s)
Miguel contrasts hydraulic and electrical actuators. Hydraulics can produce higher torque, which helps a robot recover quickly from instability and compensate for the forces created during a strong movement. Electrical actuators are cheaper and easier to power, but they are harder to control in those situations. He was impressed that Boston Dynamics' newer humanoid robot uses electrical actuators rather than hydraulics. He also observed that the new robot was not yet performing the backflips seen in earlier demonstrations, and wondered whether it would reach the same level of stability.

### Machine learning is wasteful when an exact robotics formula exists
[11:31](https://www.youtube.com/watch?v=wEmO00ADppw&t=691s)
During his PhD, Miguel tried using a neural network to calculate robot movement and solve inverse and direct kinematics. He stopped because the problem already had analytical equations. Those equations specify how actuator movements produce a trajectory for a hand or leg, while a machine learning model would only approximate the result. Miguel says deriving equations for a complex humanoid robot is difficult, which explains why some researchers use simpler robot models. He describes compressed-mass models, such as an inverted pendulum around the robot's center of gravity, alongside distributed models that represent many joints and masses.

### Sports analytics can move from tracking objects to understanding tactics
[16:18](https://www.youtube.com/watch?v=wEmO00ADppw&t=978s)
Miguel describes a football project that tracked team movements and the ball for sports analytics. Microsoft has partnered with football clubs in Europe, and he mentions work involving DeepMind and Liverpool as well as interest from Real Madrid. Computer vision can provide object detection and tracking, while data teams can study team performance and athlete health. When Demetrios asks whether this starts with technology instead of a problem, Miguel points to a corner-kick goal where a defender noticed a gap in the opposing team's strategy. He imagines AI eventually identifying similar tactical patterns and advising coaches, though he says that capability is still far away.

### Recommendation systems have batch, real-time, and hybrid designs
[26:58](https://www.youtube.com/watch?v=wEmO00ADppw&t=1618s)
Miguel describes three common architectures. A batch system trains a model, generates recommendations for each user, and stores them in a database. The front end can query that database very quickly, sometimes below 10 milliseconds, but it cannot react to what the user is doing right now. A real-time system deploys a model that scores current activity, although complex models can increase response time. A hybrid, or two-step, system stores a large candidate set and uses a second model in real time to rerank those candidates and return the most relevant items.

### Choosing a recommender is difficult because older methods remain useful
[33:06](https://www.youtube.com/watch?v=wEmO00ADppw&t=1986s)
Miguel says recommendation systems differ from areas where newer algorithms tend to replace older ones. In computer vision, he describes convolutional networks and Vision Transformers as the methods people generally use now. Recommendation teams may still choose linear models, simple models, or no machine learning, depending on their data and use case. A company with millions of songs has different needs from an e-commerce company with thousands of products. Teams must test different approaches to learn which one creates value for their users.

### Recommendation experiments consume time and often fail
[37:36](https://www.youtube.com/watch?v=wEmO00ADppw&t=2256s)
The recommendation pipeline includes data preparation, model development, deployment, and A/B testing. Miguel says an experiment may need one or two weeks before its result is statistically sound. He gives an industry average in which 80% of experiments fail. That makes recommendation work difficult to fund, especially when decision makers expect an immediate revenue increase. He argues that personalization can produce a high return, but teams need enough time and resources to test models and find an approach that works for their business.

### Leadership responds to measured financial results
[43:56](https://www.youtube.com/watch?v=wEmO00ADppw&t=2636s)
Miguel's approach to gaining support is to arrive with results rather than broad claims about AI. He tells leaders how much money his team made through its work, then waits for their response. A/B tests measure business metrics such as average revenue per user, conversion rate, or monthly active usage. The team can translate a measured improvement into an estimated dollar value using prior business results. Miguel says leaders understand a request for people and budget more easily when the result is also stated in dollars.

### Large language models can change the recommendation interface
[51:57](https://www.youtube.com/watch?v=wEmO00ADppw&t=3117s)
Miguel says large language models can be used as another algorithm inside a recommender system, alongside linear models or gradient-boosted trees. They can also power a personalized chatbot. A user could describe a product or movie through text or voice, and the system could combine that conversation with purchase history and preferences to produce recommendations. He notes that Netflix already presents recommendations through rows rather than conversation, while search experiences are moving toward chat interfaces. E-commerce search may follow the same path.

## Notable quotes
- Miguel Fierro: "If you do have the formula, why are you going to do that?" (12:20)
- Miguel Fierro: "The benefit is that from your front end you are just querying a database." (27:47)
- Miguel Fierro: "On average for most industries 80% of the time you're going to fail." (39:02)
- Miguel Fierro: "If you come with results, actually my technique is very simple." (44:04)
- Miguel Fierro: "You can think of an LLM as just another machine learning algorithm." (52:12)

## Tools & references mentioned
- Microsoft
- Boston Dynamics
- Amazon
- model predictive control
- Newton
- Lagrange
- DeepMind
- Liverpool
- Real Madrid
- Spotify
- Netflix
- TikTok
- Linux Foundation
- OpenAI
- ChatGPT
- Azure
- Microsoft Store

## Who should watch
- You are deciding whether a robotics problem needs machine learning or whether a control model and known equations are sufficient.
- You are building a recommender system and need to choose between precomputed results, real-time scoring, or candidate retrieval followed by reranking.
- You need to explain recommendation experiments and their budget to leaders in terms of measured business metrics and revenue.

## Related talks

- [Recommender System: Why They Update Models 100 Times a Day](https://mlopstalks.com/talks/recommender-system-why-they-update-models-100-times-a-day) (Gleb Abroskin, FunCorp, 49:01)
- [System Design for Recommendations and Search](https://mlopstalks.com/talks/system-design-for-recommendations-and-search) (Eugene Yan, Amazon, 58:23)
- [Product Enrichment and Recommender Systems](https://mlopstalks.com/talks/product-enrichment-and-recommender-systems) (Marc Lindner & Amr Mashlah, eezylife Inc., 56:12)
- ["Real-Time" ML: Features and Inference](https://mlopstalks.com/talks/real-time-ml-features-and-inference) (Sasha Ovsankin & Rupesh Gupta, LinkedIn, 51:55)
- [The Missing Data Stack for Physical AI](https://mlopstalks.com/talks/the-missing-data-stack-for-physical-ai) (Nikolaus West, Rerun, 52:43)