Why autonomous driving is still so hard, with Joana Fonseca | Builders

Joana Fonseca is an AI engineer at TRATON, working on AI for autonomous driving.

In this episode of Builders, Joana explains how vision language models (VLMs) and vision language action models (VLAs) could change autonomous driving by helping vehicles understand entire scenarios rather than simply detecting individual objects.

She also breaks down what makes deploying AI in the physical world so difficult, from latency and model size to safety, edge cases, testing, and the huge amounts of diverse real-world data required.

Joana holds a PhD in machine learning and robotics from KTH Royal Institute of Technology. Her work has taken her from developing algorithms for autonomous submarines to building AI systems for autonomous trucks at TRATON.In this episode:

How VLMs differ from traditional computer vision
How AI can understand an entire driving scenario
The role of VLAs in autonomous vehicles
Why 99.99% reliability may still not be enough
Why latency becomes critical when AI operates in the physical world
The data problem behind autonomous driving
Why moving from prototype to production remains so difficult
What production-ready AI looks like for autonomous systems

Subscribe to Builders for more conversations with the people building the future of technology, AI, and engineering.

Chapters

(00:00) Introduction

(01:41 )⁠ From robotics to autonomous driving

(03:49 )⁠ Building AI for autonomous trucks

(⁠04:47 )⁠ What are vision language models?

(⁠05:48 )⁠ VLMs vs traditional computer vision

(⁠07:34 )⁠ From VLMs to vision language action models

(⁠09:51 )⁠ The black box problem

(⁠12:38 )⁠ Why 99.99% safety isn’t enough

(⁠13:57 )⁠ When AI latency becomes dangerous

(⁠15:01 )⁠ Testing AI in the physical world

(⁠16:12 )⁠ The autonomous driving data problem

(⁠17:28 )⁠ Where current AI models fall short

(⁠18:49 )⁠ Why prototypes struggle to reach production

(⁠19:52 )⁠ Why robotaxis haven’t scaled everywhere

(⁠21:36 )⁠ What makes AI production-ready

Det här avsnittet är hämtat från ett öppet RSS-flöde och publiceras inte av Podme. Det kan innehålla reklam.

Avsnitt(80)

The Last AI Mile with Zimply.ai's Jesper Fredriksson | Proxify Talks

AI is rapidly evolving from a powerful tool into a transformative force for organizations and society.In this keynote, Jesper Fredriksson explores the current state of AI, the accelerating pace of tec...

27 Juli 22min

The hidden security risks of AI agents, with Ali Leylani | Proxify Talks

I agents are becoming more capable and more autonomous. But as they gain access to tools, memory, and company data, they also introduce new security risks.In this Proxify Talk, Ali Leylani, Co-founder...

24 Juli 13min

Why AI adoption fails with Louise Vanerell & Carl Carlheim-Gyllensköld | Proxify Talks

AI implementation is easy. AI adoption is hard.In this episode, Louise Vanerell and Carl Carlheim-Gyllensköld explore why so many AI initiatives fail to create lasting organizational value despite str...

15 Juli 29min

Closing the AI Gap, with Atlan AI’s Rocío Bachmaier | Proxify Talks

How do companies move beyond AI pilots and actually scale AI across their organization?In this keynote from Proxify HQ, AI strategist and transformation expert Rocío Bachmaier shares practical lessons...

8 Juli 22min

How Slack actually uses AI at work

AI is changing technical roles faster than ever. But does that mean customer-facing engineers are becoming obsolete?In this episode of Builders, Lee Haynes sits down with Liliana Lindberg, Lead Soluti...

1 Juli 47min

The end of traditional companies? How AI Is reshaping leadership, hiring & work

AI, organizational transformation, leadership, hiring, AI native companies, the future of work, organizational design, and AI adoption are changing how businesses operate.In this episode of Builders, ...

23 Juni 49min

Why most AI projects fail in production (and how Amazon solves it)

Data Engineering, Machine Learning, AI, Data Trust, AI Governance, and ML Ops are transforming how companies scale intelligent systems.In this episode of Builders, Deepak Yadav, Engineering Leader at ...

16 Juni 41min