How NVIDIA's AI Mercedes Actually Handles Berlin Traffic

What happens when NVIDIA’s AI drives a Mercedes-Benz CLA through Berlin traffic? David Bombal rides along to see how Alpamayo and the Halos safety stack work together.

At GTC Berlin, I ride in a Mercedes-Benz CLA running a development version of NVIDIA’s Alpamayo driving software. We encounter cyclists, pedestrians, narrow roads, unusual junctions, parked cars and unpredictable drivers while the car handles much of the driving under human supervision.

We look at what is actually happening behind the scenes: an end-to-end AI model proposes a driving path, a separate classical safety stack checks the environment and rules, and a trajectory arbiter decides when to allow the plan or fall back. The car uses on-board NVIDIA DRIVE Orin computing, cameras, radar and ultrasonic sensors.

We also discuss training with US and European driving data, simulation with NVIDIA Cosmos, the difference between this development vehicle and NVIDIA’s downloadable Alpamayo research models, and what it takes to adapt the system to different vehicles and countries.

In this video:
• A live AI-assisted drive through busy Berlin streets
• How Alpamayo and the classical Halos safety stack work together
• Why narrow roads, cyclists and German traffic lights are difficult cases
• The cameras, radar, ultrasonic sensors and DRIVE Orin hardware
• How real-world data and Cosmos simulation help train and test driving models
• Which Alpamayo research models you can explore yourself Important: this demonstration is a supervised Level 2 driver-assistance system.

The person in the driver’s seat remains attentive and responsible; this is not an unsupervised robotaxi. Would you trust AI-assisted driving in a busy city? Which situation would you want it to handle first?

NVIDIA Alpamayo: https://www.nvidia.com/en-us/solution…
NVIDIA Halos: https://www.nvidia.com/en-us/ai-trust…

Alpamayo research downloads: https://developer.nvidia.com/drive/do…

// TIMESTAMPS //
0:00 – Intro
0:39 – Not a normal Mercedes
01:42 – How the Alpamayo stack works
02:52 – Levels of self-driving cars
04:01 – How the cameras work
05:10 – Difference between Alpamayo and Halos
07:49 – Self-driving cars in the future
08:30 – Alpamayo vs standard LLM
10:00 – Going for a drive // Training the AI model
14:52 – Alpamayo and Halos working together
15:38 – Left side and right side driving
16:21 – Smooth driving behaviour
18:47 – Petrol (Gas) vs electric cars
19:47 – Real-world scenario driving
21:16 – Petrol (Gas) vs electric cars continued
22:56 – Unique traffic laws
25:49 – The future of Aplamayo and Halos
26:49 – Difference to other self-driving cars
28:45 – Aplamayo on Hugging Face
30:35 – Trusting self-driving cars
33:02 – Handling unusual scenarios
34:17 – Adjusting acceleration based on the vehicle
37:05 – Hardware explained
38:43 – Self-driving cars in the future pt.2
39:57 – How the model is trained
41:38 – Conclusion