Europe's first fully driverless commercial robobus runs in France.

beti is a French operator of automated mobility networks, and a spin-off from Bertolami, a family coach business that has moved people around rural France for 75 years. CEO Benjamin Baudet explains the layer of the autonomous driving industry nobody photographs: the people and software that keep a driverless bus service running once the driver is gone.

Before beti existed, the family studied its own routes and asked a simple question — if the driver is no longer on board, who does everything else? They counted 421 daily tasks that had nothing to do with driving. Autonomous driving technology replaces the driving task. It does not replace the operator. That finding became the company.

What beti sells is people and software: field teams trained to deploy someone else's autonomy, plus two in-house platforms — OperaVision for the ground crew and HyperVision for remote monitoring.

The conversation gets into the fleet operations numbers this layer rarely publishes: four vehicles per hypervisor on two-hour rotations, roughly 20% of operating effort spent on cleaning, charging, depot work and roadside recovery, and around 30% lower operating cost than a service with a driver — but only above eight hours a day and seven days a week.

In October 2024 beti flew to Guangzhou to argue it was the right size of European partner. Four months later the Valence TGV robobus line launched with WeRide and Renault. Adapting a robobus designed for the Chinese market took a year of closing gaps on language, interfaces and data protocols.

Plus: how French driverless permits came through in three months, why Keolis, Transdev and RATP are customers rather than competitors, how beti is preparing for robotaxi and goods transport alongside public transport, and the 2030 target Benjamin uses to measure success.

This show is the interview half. The analysis lands every Monday in the newsletter. Free: https://avmarketstrategist.substack.com/?utm_source=youtube&utm_medium=description&utm_campaign=ep27

TIMESTAMPS

0:00 - open: the gap in the leading pack is narrowing

0:59 - The service layer nobody photographs

1:53 - What beti actually sells

2:24 - Why a family coach operator started an autonomy company

3:32 - The 421 tasks that are not driving

5:01 - What transfers from a 75-year-old bus business

6:00 - The turnkey offer: people plus two software platforms

7:37 - Hypervision vs. teleoperation, and who owns the driving

8:52 - Why the control room sits in the same country

9:16 - Training customers to not need you

10:16 - What operators underestimate

11:12 - Four layers, and the two beti refuses to enter

12:13 - How many vehicles one hypervisor can handle

13:38 - What caps the ratio: human attention

14:18 - The ground crew: 20% of operating effort

14:54 - What stops WeRide or Renault building this themselves

15:55 - Valence 2025: how the WeRide deal came together

17:00 - Why that route, and the riders who became ambassadors

18:59 - Driverless permits in three months

19:59 - Rebuilding a Chinese robobus for Europe

22:02 - Navya, EasyMile, WeRide: a 20-criteria operator framework

23:31 - Roland-Garros 2026: a 100% beti operation

25:00 - Who the customer is, and who writes the cheque

26:11 - 30% below a driven service, with conditions

27:18 - When your supplier goes under

28:58 - Keolis, Transdev, RATP: customers, not competitors

30:13 - The 2030 target: 2,000 hypervision licences

Everyone measures robotaxi progress in miles driven. Almost nobody measures the hours the vehicle spends not driving. George Kalligeros, co-founder and CEO of Aseon Labs, is building a robot that compresses a depot into a single parking space.

When a robotaxi drops below its charge threshold, it leaves the zone it is earning in, drives three to five miles to a depot, queues, gets serviced, and drives back. In San Francisco, a compact city with unusually good infrastructure, that is roughly two hours out of service, and it happens several times a day. The driving stack is fully autonomous. Everything around it is still humans, real estate and deadhead miles.

Aseon's answer is a six-metre box that occupies one charging bay, holds two six-axis robotic arms, and completes a reset in twenty to thirty minutes: inspection, charge, interior clean, item retrieval, data sync and reboot. George is unusually direct about what version one does not do. No exterior wash. No sensor detailing. Anything genuinely unpleasant goes back to the depot.

Daniel Abreu Marques presses on the parts that decide whether this is a business or a demo: why Aseon operates a network instead of selling boxes, how the project-company and asset-backed financing model is designed to pull AV operators into owning part of their own infrastructure, and why a pod needs only about five resets a day — roughly 10% utilisation — to break even.

The wider argument is the one this show keeps returning to. If more than half the cost of a robotaxi service sits in infrastructure and operations rather than in the driver, then autonomy economics are settled off the road, not on it. And once cities start tendering licences with a cap on vehicle numbers, "drive somewhere cheaper to charge" stops being a strategy.

As George puts it: the infrastructure has to meet the vehicle, not the vehicle be taken to the infrastructure just because it drives itself.

Plus: what four years at Tier taught him about who you can actually incentivise, why Europe's density makes it the natural market for this product, and why he is building the company in California anyway.

TIMESTAMPS

0:49 - Introduction
1:26 - What Aseon is, and what an operator actually buys
1:52 - Inside the box: one bay, two robotic arms
2:41 - A 20-30 minute reset, bounded by charge rate
3:29 - What it replaces: two hours out of service, several times a day 4:50 - What they found inside real AV depots
5:31 - A third of the fleet offline at any given time
6:44 - Four years at Tier, and who you can actually incentivise
7:50 - Which jobs version one does, and which it refuses
9:47 - Interior resets: the 90% a robot can close out
10:55 - Getting power into a dense-city parking bay
12:00 - One pod, many platforms: ID Buzz down to Cybercab
13:55 - Orchestration: no queuing, just-in-time only
14:51 - When the pod fails with a vehicle inside
15:52 - Why Aseon runs a network instead of selling boxes
18:07 - How operators pay, and what makes it bankable
19:08 - Five resets a day: the utilisation math
21:02 - Where the $10M goes, and what 2027 has to prove
21:38 - $50M a metro: from venture equity to asset-backed debt
22:48 - Movable, permittable, live in eight weeks
23:55 - Lyft, Avis, Move: why route into a third-party network
25:20 - What stays in-house
26:14 - Are AV depots the new data centres?
27:39 - Is the cost really decided off the road?
28:55 - Curb space, cities and the traffic externality
30:53 - Europe: why build in the US, and when Aseon arrives here
32:52 - The one metric for the next twelve months

France already has a working legal framework for approving driverless public transport. Most of the industry has never heard of it. Aryldo Russo of GESTE Engineering takes it apart piece by piece.

Getting a vehicle type approved in Europe is one problem. Getting a driverless service running on a specific street, with a control room behind it and a city that has to agree, is a completely different one. France solved the second problem by writing its rules into the transport code rather than vehicle law, and handing the job to the national technical service that already oversaw metros and tramways.

That inheritance shapes everything. The system is assessed as a whole, not as a vehicle: six domains covering onboard safety, connectivity and positioning, cybersecurity, road infrastructure, vehicle behaviour and operations, plus a seventh that consolidates them into a single verdict. Three sequential dossiers carry a project from generic capability to a specific route. And the safety bar is inherited from rail: globally at least as safe as the service being replaced, measured against national statistics rather than one bus route.

Aryldo explains what happens when an operator calls, what counts as evidence, and why a million kilometres of driving proves less than teams expect. We also cover why the law refuses to name a vehicle-to-operator ratio for remote supervision, whether French cybersecurity rules duplicate UN R155, how you assess an end-to-end driving system you cannot decompose, where the seam sits between EU type approval and national authorisation, whether Brussels should use the French model as its template, and whether Europe's assurance layer is a safety asset or a competitive handicap against Waymo and Chinese operators.

Plus the number nobody publishes: roughly what the whole safety demonstration costs as a share of the project.

TIMESTAMPS

0:00 - Type approval is one problem. Running a service is another
1:09 - The French technical service, and its metro and tram origins
2:46 - What rail engineers see that automotive teams miss
3:45 - What actually happens when a project calls you
5:19 - The seven domains of the assessment
6:56 - Independence: assessor and engineering support at once
8:29 - What counts as an automated road transport system
9:35 - Built through the transport code, not vehicle law
10:46 - The remote intervener and the service organiser
12:02 - Predefined routes, zones, and whether Paris fits
13:36 - GAME: at least as safe as what it replaces
14:54 - National statistics, or one specific route
15:26 - Who signs: the operator, the mayor, the assessor
16:23 - Three dossiers: DCST, DPS, final safety case
19:17 - When to call an assessor, and the cost of calling late
20:45 - Remote supervision: why the law names no number
22:10 - Do remote operators have to sit in France
22:47 - French cybersecurity rules vs. UN R155
24:17 - Why a million kilometres proves very little
26:49 - Traceability when behaviour comes from data
28:26 - The seam between EU type approval and French rules
29:51 - Should Brussels use the French model
31:07 - What changes when it is a truck, not a shuttle
32:14 - How long it takes and what it costs
33:26 - Waymo, China, and whether assurance is a handicap
34:36 - The one thing he would change

LINKS

The AV Market Strategist (my weekly AV analysis) - https://avmarketstrategist.substack.com

Aryldo on LinkedIn: https://www.linkedin.com/in/aryldogrussojr/

#AutonomousDriving #Robotaxi #DriverlessVehicles #MobilityPolicy #AutonomyInsiders

WeRide CEO Tony Han on where robotaxis actually break even, how the Uber partnership works in Abu Dhabi, and why Europe could be profitable from day one.

Most robotaxi conversations cover one city, maybe one country. Tony Han runs a company that has tested or operated in more than 60 cities across 13 countries and holds autonomous driving permits in eight of them. Daniel Abreu Marques sits down with the co-founder and CEO of WeRide — the Nasdaq and Hong Kong dual-listed operator behind robotaxis, robobuses, robovans, robosweepers and a fast-growing ADAS business — to get specific about the numbers.

This is a conversation about business models, not demos. Where each product line makes money, how deployment decisions actually get made, what regional break-even means in practice, and why Tony believes almost none of his L2 competitors qualify as robotaxi companies at all.

You'll discover:

• Where WeRide is at or near break-even today, and why Europe's higher taxi fares change the math from day one
• How the three-way Abu Dhabi structure works, and who actually decides how many vehicles run on a given day
• What happened to rider behavior after Uber launched a dedicated autonomous category — and why the surprise wasn't what you'd expect
• Why more than 80% of last quarter's new vehicles went to robotaxis rather than the other four product lines
• The "qualification process" Tony proposed a year ago, and why he argues no L2 competitor has passed the reverse version of it
• Why WeRide refuses to build robotrucks or mining trucks despite the obvious adjacency
• How WeRide selects European launch cities — population, taxi density, regulator openness, and one factor most operators ignore entirely
• Why he calls the L4-to-L2 move a "projection" and the L2-to-L4 climb a "lifting" problem

Tony also gets into the parts of the story that rarely make the earnings call: the sabbatical that turned into a Baidu AI Lab role, why he left a tenured professorship for a company with no business plan, what Taoism and Bertrand Russell have to do with building autonomous systems, and why a live connection hiccup during the recording turned out to be the best possible illustration of the physical AI reliability problem.

For anyone tracking where autonomy goes next — from the Middle East build-out to Spain, Slovakia, Denmark and the open question of Germany — this is the operating picture behind the headlines.

Timestamps

0:56 - Introduction
1:58 - Leaving a tenured professorship
2:15 - Baidu AI Lab, Andrew Ng, and an unexpected desk mate
6:36 - What philosophy has to do with autonomous driving
9:05 - Five product lines, five business models
15:15 - Why WeRide won't build robotrucks
17:52 - Inside the 3,400-vehicle fleet
18:46 - How WeRide decides what gets deployed where
21:19 - Who owns the vehicles: asset-light vs. own fleet
22:32 - The Middle East doubled in one quarter
23:13 - Abu Dhabi: how the Uber partnership actually works
24:50 - What changed when riders could choose a robotaxi
26:48 - Business class at taxi prices
28:44 - Europe: Spain, Slovakia, Denmark and what comes next
30:26 - How WeRide picks a European city
32:44 - Germany, Wittgenstein and a robotaxi in Stuttgart
34:01 - Unit economics: where WeRide breaks even
36:22 - Physical AI vs. digital AI
39:03 - L4 to L2: projection, lifting and the qualification test
42:45 - What separates the long-term winners
44:31 - Closing: a live lesson in reliability

This video is sponsored by TaskUS.

Autonomous fleets do not scale on software alone. The hidden advantage is the human layer and Nick Allen reveals how TaskUs powers the remote assistance, safety, validation, and emergency response that keeps robotaxis moving when the real world gets messy.

If you think autonomy is just about removing people, this conversation flips that script. Daniel Abreu Marques sits down with Nick Allen, Head of Autonomous Vehicle Mobility and Logistics Solution at TaskUs, to unpack why the most successful AV operators are building human support into the stack from day one, not bolting it on after launch.

You'll discover:

• Why the best AV fleets treat humans as the "central nervous system" of the operation
• How remote assistance differs from remote driving, and why that distinction matters for safety and scale
• The cost math behind human-in-the-loop operations, including how teams can shrink cost per mile by 30% to 50%
• What a healthy vehicle-to-operator ratio looks like in commercial AV today, and why it changes by city, weather, and ODD
• How edge cases, emergency response, and rider support actually work in the first 60 seconds of a real incident
• Why Europe's regulatory patchwork is creating both friction and opportunity for AV expansion
  

Nick also breaks down the less visible but essential parts of the job: high-fidelity sensor labeling, scenario reconstruction, fleet management, training and recertification, resiliency support for operators, and the hybrid model TaskUs uses across customer sites and its own facilities.

We also get a look at the next wave. From Manchester to the broader UK and Europe, Nick explains why localization, GDPR, and country-by-country compliance are reshaping deployment strategy, and why AV is becoming a direct on-ramp to robotics and humanoid systems.

Find Nick here: https://www.linkedin.com/in/nicholasjallen/
TaskUs here: https://www.taskus.com/services/autonomous-vehicle-operations/

Timestamps:

01:39 TaskUs in the AV Ecosystem: From Data Annotation to Operations Coordination

03:30 Why Autonomous Vehicles Still Need Human Support

04:18 Cost per Mile and the Right Operator-to-Vehicle Ratio

05:30 When AV Companies Should Bring in Operational Support

06:36 Data Annotation and Training for Complex Edge Cases

07:47 Remote Assistance Explained: How the Command Center Works

09:01 Remote Driving vs. Remote Assistance

10:41 Latency, Safety, and Commercial Operator Ratios

13:09 Training and Qualification for Remote Assistance Teams

14:15 Own Sites or Customer-Side Operations?

15:21 Edge Cases and Emergency Response on the Road

17:47 Training, Compliance, and “Learning as a Service”

19:21 Business Models, Customers, and the Path to Commercialization

21:26 The Biggest Challenges in Scaling Autonomous Mobility

23:10 Global Teams, Quality, and Resilience in AV Operations

28:29 TaskUs’ Europe Strategy and Market Entry

31:40 Localization, Regulation, and Data Privacy in Europe

35:09 The Future of Human-in-the-Loop and Robotics

38:32 Closing

#AutonomousDriving #Robotaxi #AutonomousVehicles #RemoteAssistance #FleetOperations #AutonomyInsiders

Everyone in autonomous driving argues about sensors, foundation models and remote assistance ratios. Almost nobody argues about the last ten meters. Bern Grush, CEO of Pudocity, on why the robotaxi industry's real constraint is not the AI. It is deciding where the vehicle stops.

Safety is an engineering problem, and engineering problems eventually get solved. Pickup and drop-off is not. It is a spatial management problem, and every robotaxi competes for a fixed resource at the curb against every other vehicle in the city. Which produces the prediction at the center of this conversation: as the industry scales, it will get safer and more congested at the same time.

A fleet of 1,000 robotaxis generates roughly 25 million pickup, drop-off and staging events per year. Even a small failure rate becomes a million annual opportunities for a blocked bike lane, a blocked bus lane or a blocked fire truck. We cover what actually goes wrong at the curb, why better AI cannot fix a conflict that is fundamentally social, why five vehicles at one address can break a street, and what happens when three competing fleets surge to the same stadium curb with no coordination layer.

We also get into how Pudocity's Curb Reserver works, who licenses it, who pays and who owns the data, why Bern drafts the ISO standards and builds the product at the same time, and the part that matters most for Europe: US states have preempted their own cities out of curb management, EU urban density is three to four times higher, and the window to write one clause into deployment permits is open right now.

TIMESTAMPS

0:00 - Safer and more congested at the same time
0:55 - Nobody is debating the last ten meters
1:52 - The claim: the Achilles heel is not the AI
3:40 - But Waymo does hundreds of thousands of rides a week
5:11 - Why AI cannot replicate human improvisation at the curb
6:19 - Hard problem, or just an unglamorous one
8:16 - What actually goes wrong at the curb today
10:04 - When annoyance becomes a systemic problem
11:09 - Vans, buses, bikes: new problem or exposed problem
12:43 - Accessibility, and why it is also a commercial issue
14:05 - Public space for private fleets: reservation is not ownership 15:32 - Three fleets, one stadium, no coordination layer
17:04 - Inside Pudocity: how the Curb Reserver works
18:33 - Why the orchestration layer has to be independent
19:56 - Who actually pays for it
21:06 - Beyond robotaxis: goods, charging, staging, personal AVs
23:11 - Selling to cities vs. operators, and the biggest surprise
25:13 - The moat: patents, standards, network effects
26:41 - ISO 4448 and ISO 25614 without the jargon
29:24 - Why standards, not lobbying, are the leverage point
30:33 - Do standards freeze immature technology too early
31:59 - Inside the WEF Global Future Council on Autonomous Systems 34:13 - Is the curb problem worse in Barcelona than in Phoenix
35:45 - The one clause every transport minister should write into permits
39:02 - Who should own the curb data
39:58 - The AV belief Bern says is simply wrong
41:42 - Next steps for operators, cities and investors
43:48 - Closing thoughts

LINKS

The AV Market Strategist (newsletter) - https://avmarketstrategist.substack.com

Europe's autonomous vehicle problem is not the technology. It is that almost nobody outside the industry knows what is coming. Guido Di Pasquale, Managing Director of PAVE Europe, on trust, education, and what actually decides whether Europe becomes a player or a market.

Most AV coverage tracks the technology, the funding rounds and the deployment maps. This conversation covers the layer underneath all of it: whether the European public and European policymakers are anywhere near ready for what is being built. Guido Di Pasquale runs PAVE Europe, the nonprofit coalition Partners for Automated Vehicle Education, whose members span AV developers, insurers, certification bodies, public transport operators and public authorities. He has spent more than a decade inside the European debate on autonomous mobility.

We get into the uncomfortable structural questions first: how a coalition funded by the industry stays a neutral educator, and why PAVE Europe's founding mix of insurers and testing organizations, rather than carmakers, tells you something real about how Europe approaches autonomy. From there we cover what the survey data actually says about European trust, why the European concern is governance and jobs rather than crash safety, and the chicken and egg problem of building trust in a continent with almost nothing to try.

We also cover accessibility as a design requirement rather than an afterthought, why Europe's collaboration framework is a genuine and defensible edge, what Europe is really missing beyond capital, whether Europe should restrict Chinese AV players the way Washington has, and what a realistic 2030 looks like

TIMESTAMPS

0:00 - Why public readiness is the AV story that gets no airtime
0:30 - What PAVE Europe is, and what it deliberately is not
2:20 - The neutrality question: the members pay the bills
3:58 - What "education" actually looks like month to month
6:52 - Why PAVE Europe was founded by insurers and TÜV, not carmakers
8:26 - Policymakers now, citizens later: where the energy goes
10:36 - The study visit that changed a national regulation
12:51 - What Europeans actually fear, and what they get wrong
15:45 - The chicken and egg problem: trust requires trying
17:46 - Accessibility: repairing the mistakes of the old mobility system 20:43 - Is European acceptance uniform? Exposure vs. acceptance 22:28 - Europe's real and defensible edge
24:14 - The single biggest thing Europe is missing
26:16 - The talent Europe educates and then exports
26:47 - Regulate first vs. deploy first: strength or excuse?
28:31 - What Europe can learn from China, and what it cannot copy 30:01 - Should Europe restrict Chinese AV players?
32:09 - Setting standards instead of building walls
32:52 - Europe in 2030: what meaningful deployment looks like
35:09 - The missing players: where are the OEMs and Tier 1s?
35:23 - Closing thoughts

Connect

• Newsletter: https://avmarketstrategist.substack.com
• LinkedIn: https://www.linkedin.com/in/daniel-abreu-marques-/
  

Certification, type approval, homologation. Not the sexiest words in autonomous driving, but arguably the most decisive ones for whether AVs ever reach European roads at scale.

The EU's Level 4 regulation (2022/1426) has existed for four years. As of today, not a single vehicle holds EU type approval under it. The common narrative blames slow regulation. My guest tells a different story: the framework is ready, the technical services are ready, and the real bottleneck sits inside the companies themselves.

Dirk Fratzke is Project Leader in the Autonomous Driving & ADAS department at TÜV SÜD, the independent testing organization that assesses AV companies across three continents, from Waymo and Aurora in the US to projects across Europe and Asia. He led TÜV SÜD's involvement in the EVA shuttle project in Karlsruhe, one of Germany's first SAE Level 4 deployments in public transport, and has co-authored academic research on systematically testing autonomous driving functions.

In this episode, we cover:

• What homologation and type approval actually mean, and how Europe's independent assessment model differs from US self-certification
  
• The two-step approval process for Level 4: EU type approval for the product, then operating area approval from the "owner of the road"
  
• Why the 1,400-unit small series cap is a benefit for AV companies, not a showstopper
  
• What is really delaying Europe's first Level 4 approval, and why Dirk says "maybe one year, maybe two"
  
• Why Mobileye's completed functional safety management assessment matters as a milestone
  
• Whether end-to-end AI systems can pass European certification at all
  
• Why Pony.ai, WeRide, and Momenta are entering Europe
  
• How TÜV SÜD approved shuttles without steering wheels by arguing from the regulatory framework for horse-drawn coaches
  
• The newly adopted UN global regulation for driverless vehicles, and the alcohol interlock absurdity that shows why updates were overdue
  

Whether you build, operate, insure, or invest in autonomous vehicles, this episode explains the process that decides who gets to deploy in Europe and when.

Timestamps:

00:00 – Cold open: why AVs need an independent set of eyes
01:18 – Intro: certification, type approval, homologation
03:07 – Dirk's path: from car mechanic to TÜV SÜD
04:48 – Why a 160-year-old testing org jumped into autonomous driving
06:19 – What a "technical service" actually does
07:22 – Team size and how demand has shifted
09:24 – Homologation vs type approval vs US self-certification
11:10 – Level 3 vs Level 4: two separate regulatory tracks
14:00 – How UNECE regulation becomes binding law
17:05 – The EVA shuttle project in Karlsruhe (and Bad Birnbach first) 19:05 – Approving a vehicle with no steering wheel
21:09 – The two-step EU approval process (2022/1426)
23:30 – Operating area approval: the "owner of the road"
25:11 – Zero Level 4 approvals: what's really blocking it
27:32 – Vertical integration vs partnership: who applies?
30:21 – How long approval realistically takes
32:43 – The 1,400-unit small series cap, and why it's a benefit
35:12 – Why US and Chinese companies come to TÜV SÜD
36:46 – Waymo, self-certification, and the four-eye principle
39:15 – Chinese AV companies entering Europe: the Level 3 play
41:42 – Can end-to-end AI pass European certification?
44:16 – The new UN global regulation for driverless vehicles
47:00 – When can a passenger hail a driverless AV in Europe?
49:00 – The one thing Dirk would change tomorrow

Most conversations about autonomous driving stop at the vehicle: the sensors, the AI, the next demo. But once an autonomous service carries real passengers in public transport, the hard part is not the car. It is everything around it.

In Episode 19 of Autonomy Insiders, I sit down with Stefanie Berliner, who leads the DACH business at Holo, and Anders Toft, who runs operations. Holo operates autonomous vehicles for Ruter in Oslo and enables operators to run autonomous mobility services across Europe.

We map who actually does what in a real AV deployment, and why the permit holder, a role most people have never heard of, carries the liability when something goes wrong. Anders walks through the first minutes after an incident, including the critical call between a local issue and a global one that grounds the entire fleet. Stefanie explains why the industry has shifted from proving technology to proving operational scalability, and why cities asking for "just one showcase vehicle" are getting turned away.

We also get into the metrics that matter: why autonomy per road type is more honest than autonomy per kilometer, why remote assistance decisions need to happen within two to five seconds, and what a financially viable operator-to-vehicle ratio looks like.

Topics covered:

• The shift from proving technology to proving operational scalability
• Who does what in a deployment: permit holder, operator, PTA, OEM, software provider
• Why the permit holder, not the vehicle maker, is liable when something goes wrong
• Incident management: root cause analysis, local vs. global issues, grounding the fleet
• What regulators in Norway and Switzerland actually require, and how differently they work
• Why autonomy per road type reveals what total mileage hides
• Remote assistance thresholds and operator-to-vehicle ratios
• How operational experience transfers between platforms and compresses deployment timelines
• What operators must build before Level 4 arrives in Europe
  
  Timestamps:
  00:00 Intro
  01:40 What Holo Does & Where It Sits in the AV Ecosystem
  02:27 Current Deployments: From Pilots to Passenger Operations
  03:40 Why the Car Is the Easy Part: Proving Operational Scalability
  06:39 What Organizations Underestimate Before a Project Starts
  08:21 The Gap Between a Pilot and Continuous Public Service
  09:55 "We Want AVs Next Year": What Cities Must Bring to the Table
  11:30 Who Does What: Mapping the Players in a Real Deployment
  14:21 Contracts, Liability Split & the Vehicle "Black Box"
  15:37 Where Dependencies Create the Most Operational Risk
  17:11 When Something Goes Wrong: Root Cause & Owning the Answer
  19:01 The Permit Holder Explained & the Setup With Ruter
  21:46 What Regulators Want to See: Norway vs. Switzerland, Data & Traceability
  27:51 Safety as an Organizational Task, Not Just a Technical One
  30:23 Inside an AV Safety Management System
  31:56 The First Minutes After an Incident: Local vs. Global
  34:38 Autonomy per Road Type: Why the Simple Number Misleads
  37:39 Metrics of a Maturing Deployment vs. Just Accumulating Miles
  39:09 Remote Assistance: Thresholds & the Operator-to-Vehicle Ratio
  41:59 SLAs When Software Behavior Changes Every Month
  44:26 Edge Cases: How Rare Events in One City Improve Another
  46:08 Where Operators and Tech Providers Misunderstand Each Other
  47:59 What 10,000+ Hours on Mobileye Transferred to MOIA
  52:05 The Operator's Role as EU Regulation Matures
  54:15 Wrap-Up

The autonomous driving industry has mostly converged on one architecture: a single, monolithic end-to-end model trained to handle the entire driving task.

Igal Raichelgauz, founder and CEO of Autobrains, argues that approach hits a wall on edge cases and compute cost.

His alternative is "agentic AI." Driving is broken into a large but finite set of specialized agents, each an expert in a specific situation, invoked by an orchestrator in real time. Autobrains claims this runs on roughly an order of magnitude less compute, needs almost no labeling, stays vision-only, and skips HD maps.

We recorded just after Autobrains' two announcements at NVIDIA GTC Taipei: a Munich robotaxi program with Uber on NVIDIA DRIVE Hyperion, and a Level 4 robotaxi platform with VinFast for Southeast Asia's high-density traffic. We get into why Munich is a deliberate choice, what "OEM-agnostic" actually buys an automaker, and how an investor list that doubles as a customer list (BMW, Toyota, Continental, VinFast, Magna, Knorr-Bremse) shapes the company.

Timestamps:

• (01:30) What Autobrains does, and the problem of bringing autonomy to 1.5 billion cars
• (02:10) Agentic AI in one sentence, and why specialization beats one big model
• (05:04) Why Igal thinks the end-to-end consensus is wrong
• (07:16) Edge cases: school buses, pedestrian intent, a kid chasing a ball
• (08:19) The claim of ~10x less compute and near-zero labeling, and how to prove it to a skeptical OEM
• (09:31) Traceability and explainability for regulators through modularity
• (11:06) The business model: mass-production ADAS, robotaxi, and air-to-road redundancy
• (13:40) Software-only, chip-agnostic, vision-first, and what that trade-off costs
• (16:12) The investor list that doubles as the customer list
• (18:40) The Munich robotaxi program with Uber and NVIDIA, and what OEM-agnostic means
• (24:05) The VinFast deal for Southeast Asia
• (29:09) Why start in Munich, one of Europe's hardest cities
• (30:51) Is Europe really behind the US and China?
• (31:36) The regulatory path from pilot to paid service, and the 50,000-hour intervention target
• (33:28) A 5-year European robotaxi outlook
• (34:34) The eyes-off / Level 3 debate and why planned handover matters

LINKS

My free weekly Newsletter: https://avmarketstrategist.substack.com/

My LinkedIn feel free to connect: https://www.linkedin.com/in/daniel-abreu-marques-/