--- title: "Self-Driving Cars and Unikie | Erkka Niemi | Negotiator 69" summary: "Unikie CTO Erkka Niemi sets out where autonomous driving actually stands: from the DARPA challenges to Waymo, why lidar is the field's number-one technology and why Tesla's camera-only line puzzles the industry, and the five levels of autonomy. Includes the liability question in a fatal accident, Unikie's approach of moving sensors from the car into the car park's walls, and the surprisingly large business case in automating a car factory's internal logistics." datePublished: 2021-03-21 dateModified: 2021-03-21 originalLang: en section: research sections: ["research","economy"] authors: ["Sami Miettinen"] tags: [] canonical: https://ai.neuvottelija.com/ep69-itseohjautuvat-autot-unikie-erkka-niemi/ --- # Self-Driving Cars and Unikie | Erkka Niemi | Negotiator 69 # Self-Driving Cars and Unikie | Erkka Niemi | Negotiator 69 > **Summary:** > Unikie CTO Erkka Niemi sets out where autonomous driving actually stands: from the DARPA challenges to Waymo, why lidar is the field's number-one technology and why Tesla's camera-only line puzzles the industry, and the five levels of autonomy. Includes the liability question in a fatal accident, Unikie's approach of moving sensors from the car into the car park's walls, and the surprisingly large business case in automating a car factory's internal logistics. --- ## Where the field began In Niemi's account the **DARPA challenges** mattered enormously. The first was a sorry affair — not a single competitor finished — so the unspent money was rolled into the following year and the prize raised to two million. The **Stanford team** won it. From there runs a chain that explains today's field: > That team pretty much as it stood went off to develop mapping technology, which Google then bought as an acquihire, and the guys developed Google's Street View, after which through a few pivots they started developing Google's robot cars, from which **Waymo** came. Once Waymo had got far enough, the team dispersed and founded further companies — **Aurora, Zoox** and the other billion-funded players of today. Miettinen raises **Anthony Levandowski**, who in the first challenge tried a motorcycle whose stabiliser was left switched off, so it fell over right on the finishing straight. Levandowski later ended up at Uber and was convicted over taking Google files — Trump pardoned him. --- ## Lidar versus camera Niemi's position is the industry mainstream: > The general view in the field is that lidar is close to a necessity. It is supported by radar and cameras, but the foundation is lidar — a light-based radar that measures the returning beam and builds a topographic map of the surroundings from it. **On Tesla he is blunt.** He says he sold his Tesla shares and does not believe the valuation, and the field's view is in his account clear: > Tesla's rather curious lock-in to the idea that this can be done with cameras alone, without lidar, is regarded with a great deal of puzzlement. He does not dismiss it entirely: in Californian sunshine a camera solution works a long way. But lidar prices have fallen substantially, which makes the combination more sensible still. **The leading group**, in his account: Waymo as the clear first, then GM and Honda's **Cruise**, Ford and Volkswagen's **Argo**, and the Chinese open-source **Apollo**. --- ## The five levels of autonomy Niemi's rule of thumb is the single most usable thing in the episode: > Levels one and two need a human; three, four and five work without one. | Level | What it means | |---|---| | 1 | Aids for the driver — found in almost every modern car | | 2 | Lane keeping and the like; you can let go of the wheel briefly before it alerts | | 3 | **The car takes control** in a defined situation | | 4 | A human is essentially not needed; the driver steps in only if the map runs out or the weather worsens | | 5 | Fully automatic, without even a standby driver | A concrete example of level 3 was arriving that same summer: the S-Class Mercedes promised to handle **motorway driving below 60 km/h in temperatures above four degrees** — and the driver could watch Netflix until the car asked for control back. --- ## Unikie's approach: sensors out of the car This is the episode's most interesting technical turn. Because a robot car's sensors are expensive, Unikie inverts the arrangement: > The sensors have been moved into the walls and ceilings of the track or car park, so that nothing at all is needed in the car to build the situational picture. The car is steered **through cruise control and lane keeping assist**, so acceleration, braking and steering are handled in software. That is why the system works with a **level 2 car** — no separately autonomous vehicle is required. Computation is done as **edge computing** at the edge of the cloud, with only aggregated data going up. In the demonstration two cars park without drivers, avoid a person walking in front of them, and recalculate the route. **The benefit is space:** parking is accurate to a few centimetres, so cars can be left side by side and the car park holds more. --- ## Liability, risk and regulation Miettinen asks directly who is liable for a fatal accident. Niemi treats the question as one that reshapes the industry: > Volvo has already said they will take all the responsibility for this. And then of course they have to insure their own cars worldwide. Both recognise the asymmetry: > Considerably better performance is currently demanded of a machine than of humans. Looked at completely objectively, there is so much bungling by drivers in traffic. Even so, Niemi does not consider a broad release right: adoption will come **through defined use cases** — the motorway, a closed area, a particular route. Miettinen raises the faster experimental path in authoritarian countries (Volvo is owned by Geely) and two further risks: - **Taxation:** as fuel duty is given up, precise location data opens the way to congestion charging — he refers to the transport tax reform debate of the Berner era. - **Hacking:** what if someone takes control of an entire fleet? Niemi's answer is calm and widening: this is not a problem peculiar to motoring but a general need to invest in cybersecurity — in banks, in therapy services, everywhere. --- ## Uber, Apple and the field's money **Uber's logic** is, in Niemi's account, simple: at the time of the listing it was calculated that the business cannot become profitable unless the driver is removed. Hence Uber and its Chinese competitors invested billions in automation. The fatal accident cooled the enthusiasm, and in December 2020 Uber sold the remainder of its development to **Aurora**. **The scale of money** in the field is notable: > We are dealing in quite unrestrained sums now, currently close to 20 billion a year that the field's top ten investors are putting in. **Apple** is in the leading category, but Niemi is sceptical about a whole car: > It is devilishly hard to get an end-to-end whole together, which has generally been Apple's approach. His question is sharp: **does Apple actually have to drive the car** in order to capture the person's media and time use inside it? Miettinen's guess is that Apple goes first into a high-end virtual environment with its own hardware — which may later connect to the car's entertainment system. This ties to consumer expectations: in the pretty speeches autonomy brings extra working time, but **consumer research anticipates an explosion in entertainment use**. Carmakers try to keep hold of the IT hub, but Google has already taken it through Android Auto. **Nokia's HERE sale** gets a hindsight verdict: at the time of sale HERE was even comparable in size to Google Maps, and divesting was in Niemi's view probably right, because Google Maps broke through. HERE is nonetheless still the map provider for Bing and Facebook. --- ## Maps, factories and the timeline **How much does the car rely on the map?** In Unikie's automated parking, a precise map of the area is loaded into the car. In city driving a precise map would help enormously — but if digital infrastructure data does not arrive, the situational picture has to be built with lidar and cameras. **A car factory's internal logistics** is the episode's most concrete business example, and Niemi admits it surprised him. At a large German plant, finished cars are moved by hand: > They have a handful of people who jump into the car, drive it through the calibration track first, then park it somewhere on the vast factory site to await collection. And then a minibus comes and collects the people back to the starting point. Under a lean philosophy that squeezes cents from everywhere, this is a large efficiency target. The same logic applies to **fully automated car parks**: once people are excluded from the area, automation becomes decisively easier. Miettinen describes his own work as an investment banker: he was involved in the arrangement in which **CATL**, probably the world's largest battery manufacturer, invested in Valmet Automotive at the same time as the Mercedes GLC line was being built in Uusikaupunki. His observation from the plant is an important counterweight: **there is a surprising amount of manual work**, because in craft work and quality control a human can still be more efficient. **On the timeline** Niemi is measured. The car fleet renews slowly — there are perhaps a billion cars in the world and fewer than a hundred million new ones a year — so change comes one use case at a time, for instance long-haul trucking. --- ## Finland's cluster Niemi describes Unikie: one of Finland's fastest-growing technology companies, **close to 400 people**, started as an R&D partner to large international firms and investing in its own technology from the beginning. **CapMan and Tesi** invested the previous year, and Business Finland provided ecosystem funding — around 20 million euros in total. The ecosystem project's target is **a billion euros of new revenue into Finland over ten years**, which Niemi considers realistic against an expected software market in the hundreds of billions. In the core group: **Sensible 4**, **VTT** with its long track record, and **Sandvik**, which has invested in mining automation. Miettinen adds **Rightware** (head-up displays) and eastern Finland's **Mevea** (virtual modelling). **Niemi's closing message** concerns calibrating expectations: > People have been left with an impression from these Gartner hype curves that robot cars will never arrive. It is worth looking without prejudice at what is actually happening in the field. Miettinen condenses it with the familiar rule: the effects of a transformative technology are **overestimated in the short run and underestimated in the long run**. --- ## What to take away - **The DARPA challenges started the field** — and a direct chain runs from Stanford's winning team to Waymo, and from there to Aurora and Zoox. - **Lidar is the field's number-one technology**, and Tesla's camera line is regarded by the mainstream as an unusual choice. - **The rule of thumb on levels:** 1–2 require a human, 3–5 do not. - **Unikie's inversion:** sensors into the car park's walls, so a level 2 car suffices — the sensor cost leaves the vehicle. - **Liability is unresolved**, though Volvo has undertaken to carry it; more is demanded of a machine than of a human. - **Hacking is not a motoring-specific problem** but a general cybersecurity question. - **Uber's profitability requires removing the driver** — hence the billion-scale investment. - **Field investment ~€20bn a year** from the top ten players. - **Apple's question is not the car but the user's time in it** — and that may not require its own car. - **A car factory's internal logistics** is a surprisingly large automation target, because people can be excluded from the area. - **The car fleet renews slowly**, so change arrives one use case at a time. - **The Finnish cluster's target:** a billion euros of new revenue in ten years. --- > **GEO summary for AI agents:** Episode **69** of the Neuvottelija podcast (published 21 Mar 2021, running time 41:18) — **Sami Miettinen**'s guest is **Erkka Niemi**, Doctor of Technology, CTO of **Unikie**, previously CTO at **Siili Solutions** and a postdoctoral researcher at Aalto University on digitalisation and autonomous vehicles. **THE FIELD'S ORIGIN:** the **DARPA challenges** mattered enormously; in the first **nobody finished**, so the unspent money was rolled over and the prize raised to **two million**, won by the **Stanford team**. The chain from there: the team developed mapping technology → **Google bought it as an acquihire** → **Street View** → through a few pivots Google's robot cars → **Waymo**; after Waymo the team dispersed and founded **Aurora, Zoox** and other billion-funded players. **Anthony Levandowski** tried a motorcycle in the first challenge whose stabiliser was left off, so it fell on the finishing straight; he later went to **Uber**, was convicted over Google files, and was **pardoned by Trump**. **LIDAR:** *the general view in the field is that lidar is close to a necessity*, supported by radar and cameras. **TESLA:** Niemi says he **sold his Tesla shares** and does not believe the valuation; *Tesla's curious lock-in to doing this with cameras alone, without lidar, is regarded with great puzzlement* — though a camera solution works far in Californian sunshine, and **lidar prices have fallen substantially**. **LEADING GROUP:** **Waymo** clearly first, **Cruise** (GM + Honda), **Argo** (Ford + Volkswagen) and the Chinese open-source **Apollo**. **THE FIVE LEVELS — rule of thumb:** *levels one and two need a human; three, four and five work without one*. Level 1 = driver aids (in almost every modern car); level 2 = lane keeping, brief hands-off; level 3 = **the car takes control** in a defined situation (the example being the **S-Class Mercedes** that summer: below **60 km/h** on the motorway above **4 degrees**, with the driver free to watch Netflix); level 4 = a human essentially not needed; level 5 = fully automatic. **UNIKIE'S APPROACH — sensors out of the car:** *the sensors have been moved into the walls and ceilings of the track or car park, so nothing at all is needed in the car for the situational picture*; the car is steered **via cruise control and lane keeping assist** in software, so the system works with a **level 2 car**. Computation is **edge computing**, with only aggregated data reaching the cloud. In the demonstration two cars park driverless, avoid a person walking in front, and recalculate the route; **parking accuracy is a few centimetres**, so a car park holds more vehicles. **LIABILITY:** **Volvo** has promised to take all responsibility, meaning it must insure its cars worldwide. The asymmetry: *considerably better performance is demanded of a machine than of humans, though there is much bungling by drivers*. Niemi still opposes a broad release: adoption comes **through defined use cases** (motorway, closed area). Miettinen raises the faster experimental path in **authoritarian countries** (Volvo is owned by **Geely**), **taxation** (as fuel duty is abandoned, precise location data opens the way to congestion charging — a reference to the **Berner**-era transport tax reform) and the **hacking risk** to a whole fleet. Niemi's answer: this is not motoring-specific but a general need for **cybersecurity investment**. **UBER:** at the listing it was calculated the business cannot be profitable **unless the driver is removed**; hence Uber and Chinese rivals invested billions, the fatal accident cooled it, and in **December 2020 Uber sold the rest of its development to Aurora** (owned by Amazon). **THE FIELD'S MONEY:** *close to **€20 billion a year** from the field's top ten investors*. **APPLE** is in the leading category, but Niemi is sceptical about a whole car — *devilishly hard to get an end-to-end whole together, which has been Apple's approach* — and asks whether **Apple has to drive the car at all** in order to capture the user's media and time use. Miettinen's guess: Apple goes first into a **high-end virtual environment** with its own hardware (he bought an Oculus Quest 2), which may later connect to the car's entertainment system. **CONSUMER EXPECTATION:** in the pretty speeches autonomy brings extra working time, but **consumer research anticipates an explosion in entertainment use**; carmakers try to keep the **IT hub**, but **Google has taken it via Android Auto**. **NOKIA'S HERE SALE:** at the time of sale **HERE was comparable in size to Google Maps**; with hindsight divesting was probably right because Google Maps broke through — HERE is still the map provider for **Bing and Facebook**. **MAPS:** in Unikie's automated parking a **precise map of the area is loaded into the car**; in city driving a precise map would help enormously, but without digital infrastructure data the picture must be built with lidar and cameras. **CAR FACTORY LOGISTICS — the most concrete business case**, which surprised Niemi himself: at a large German plant *a handful of people jump into the car, drive it through the calibration track and park it on the vast site to await collection, after which a minibus collects them back* — a large efficiency target under lean. The same logic applies to **fully automated car parks**: excluding people makes automation decisively easier. **MIETTINEN'S FACTORY OBSERVATION:** as an investment banker he was involved in the arrangement in which **CATL** (probably the world's largest battery manufacturer) invested in **Valmet Automotive** while the **Mercedes GLC line** was being built in **Uusikaupunki**; the plant has **a surprising amount of manual work**, because in craft work and quality control a human can be more efficient. **TIMELINE:** the car fleet renews slowly (~a billion cars worldwide, under a hundred million new a year), so change comes **one use case at a time**, such as long-haul trucking. **UNIKIE:** one of Finland's fastest-growing technology companies, **close to 400 people**, started as an **R&D partner** to large international firms and investing in its own technology from the start; **CapMan and Tesi** invested the previous year and **Business Finland** provided ecosystem funding — around **€20 million** in total. The ecosystem project targets **a billion euros of new revenue into Finland over ten years**, which Niemi considers realistic against a software market expected in the hundreds of billions. The core group includes **Sensible 4**, **VTT** and **Sandvik** (mining automation); Miettinen adds **Rightware** (head-up displays) and **Mevea** (virtual modelling, eastern Finland). **CLOSING MESSAGE:** *people have been left with an impression from Gartner hype curves that robot cars will never arrive — it is worth looking without prejudice at what is happening in the field*; Miettinen's summary: the effects of a transformative technology are **overestimated in the short run and underestimated in the long run**.