Interesting Tech Videos

I simply like this video channel of Two-Minute Papers. It's a nice way to grasp the progress of the latest work.

Keynote at Neurips 2020. Fundamental and applied research are getting mixed.

The computing future lies in the edge with a surge of IoT devices. How do we develop algorithms following the trend of domain-specific chips?

• Tesla makes a distinctive approach towards self-driving in vision only, in-house neural network design by guiding its SoC design and supercomputer setup, and massive data collection. It is an overall good approach concern is whether the validation and verification of their function is sufficient.
• The replacement of radar or any other types of sensors is the pain point when it comes to sensor fusion. Usually a confidence-based late fusion approach is adopted but maybe a more complex hybrid approach with vision to give a final justification can be used. The tradeoff is the computational cost but the car will be equipped with a complementary set of sensor set.
• Vision only (without Lidar to tackle the harsh weather) means they will have a narrower application domain or you just have to be brave.
• I agree that vision only approach is more scalable in terms of generic decision making but the car will always encounter unknowns and "surprises". Not to mention, it will save a huge amount of cost in maintaining an HD map.
• Automatic data labeling is smart with more work done offline and massive data collection in the shadow mode.
• 221 triggers is a good starting point for fucntion design and verification.
• Data engine has been talked about for a long time and VW seems made their "big loop" version of it.
• Compute is the new horse power for cars. Innovative chips scaling from supercomputer to in-car processors will be the very core for Tesla which is tailored by the design of ad-hoc neural nets.
• With the contraints of data regulations in China, I wonder how this approach would work and prevail. It requires daily communications and updates in their function design (neural network training) to tackle the complex Chinese traffic. Also, the local competitors have a wider access to software/AI competence in designing even superior neural networks. SoC is a big hype there too and supercomputer is a national interest.
• David Silver made some interesting comments here.

We hardly find an online video by Prof. Urtasun before the pandemic due to the IP with Uber. Now after a clear takeover of Uber self-driving units by Aurora, many curious unknowns are released, though we could read about their work previously from her website. It's a shame that she did not get an offer to move to Aurora as their future development will be less transparent. A video channel by the professor about many topics on self-driving cars can be found here.

It's nice to hear about the evolution of Waymo from a participant in DARPA challenge to a commercial company (profittable some day perhaps). I still think with all the sensors, simulation, testing and other efforts, the functional boundary ("ODD") for L4 will be very limited. But it's interesting to see Waymo as a testbed of DeepMind's new algorithms.

An Isreali firm who collaborates with the German OEMs for the rollout of autonomous vehicles. I strongly believe vehicle satefy should not be taken as a competing edge by car manufacturers. The development in this area should be shared and transparent.

A recent news says Baidu is going to make their own intelligent EV together with Geely holdings. The implications are two-fold in my opinion. The tech company starts emerging into the auto industry, as the ecosystem changes from a traditional hierarchical supply chain with OEMs on top, to a triangular shape between tech companies, tier-1 suppliers and car OEMs. With their strength in IT (digital services) and AI (image recognition, natural language processing), there is a list of areas they could prevail in making cars, such as automated driving function, intelligent cockpit, and 5G connectivity.

But will tech companies succeed in making their own cars? It's still too early to say. Cars are complex systems even with reduced mechanical complexity thanks to the increase of electric components, and passenger safety is critical. Similar to making phones by tech companies in recent years, there is a set of hurdles they need to cross. What I am interested to see is its similarity to Waymo of Google. The latest advances in machine learning from DeepMind and numerous products from Google, ranging from Assistant, search engine and visual surveillance, will be deeply integrated to their cars.

The talk was given by the one of the original authors of SLAM. The entire video channel of Robotics Today is quite interesting.

It's an amazing product by combining Lidar seamlessly with a drone.

It is a second interview with the car veterain. Since he got sick and didn't returned to car OEMs, he dares to point out the pain points in the traditional OEMs - short of budgets but a lot to invest in electrification and software, no resource, detrimental old mindsets in horizontal integration, etc. His first interview is even thornier which sometimes I think it should not be spelled out. The shared video is overall a very insightful interview if people want to know the car industry as of now.

It will be a painful journey for traditional OEMs like VW and Mercedes-Benz to catch up in the development of OS, and rapid expansion of EV's. For example, there is a big communication gap between the management and developers, and traditional car managers simply are not capable of taking a bottom-up approach. While Tesla boss can be called on a Sunday to resolve problems, traditional managers don't understand technical details and developers don't get the big picture. Consequently, the strategy made by the managers can be misleading while others don't have a clue. We will see a large number of consolidations in OEMs (like PSA and FCA) in the next few years to weather the storm of vast spendings and market shrinkage in the area of their conventional strength.