MIT self driving learning note

All the videos and slides could be found here
or in bilibili

1. Lecture Introduction

Prof from MIT gives us an overview of the self driving techniques.

1.1 Requirements for automony vehicles.

He offered us an opinion that I found very intriguing.

• automony vehicles will not be adopted because they are safer.
• automony vehicles will not be adopted because they are cheaper.
• automony vehicles will not be adopted because they are faster.
• automony vehicles will be adopted because they offered a better human experience.

In his opinion, if we want to make the techniques being wieldly used, we need to offer a good human experience. HE was not telling that the other factors are not important. In fact, they are essential. But they are not enough for every ordinary people to accept it as a daily tool.

1.2 Level of automony

There is a world wide spreaded standard of the level of automony, from L0 to L5 a fully automony.

However, the prof indicated that they can actually be divided into two categories.

• Human interfered automony. which involves human to be a necessary part of the system.
• Fully automony. which is independent of human, but still need to allow human involvement.

There are companies focus on L2-L3 a human interface structure. For an example, Tesla. And other companies, directly aiming at fully automony. For an example, Waymo.

1.3 Popular views on automatic vechicles

• Driving task is easy -> n , it is not easy
• Humans are bad at driving ->Human is good at controlling
• Human and antomatic cars won’t mix well -> Actually 34.4% of the miles driven on plane are with autopilot engaged

1.4 Two popular approaches

• Vison Sensors + Deep learning
• Lidar + HD Maps

These sensors have different advantages.

2. Sterling Anderson (Aurora)

(Co-Founder Aurora)

2.1 Co-pilot System

He talked about the co-working system of human and algorithm to make a trajectory. the algorithms is based on optimization theory.

• Give objective and some constraints (e.g. keep in the road). It will work out with a trajectory.
• And the system will adjust human’s action.
• Human experience of this actually depends on the driver. Lots of human interface challenges.

2.2 Driving match Experiment

Experiment:

• the Co-ploter make the journey more save and faster.
• The driver feels more control over the car(12%larger) when the Co-piolet involved. (the vehicle is what I want it to do, but not what I told it to do)
  

2.3 Some questions

• economic bottleneck: sensor cost
• technical bottleneck: main challenge is predict the behavior of other actors.

He sees Lidar sensor a method for fast deployment of the self-driving system. Lidar is not a necessary part of the system. It also depends on the market target.

3. Drago Anguelov (Waymo)

3.1 System start of art

He firstly talked about the challanges of the essential parts of the system :

Preception (sensor to scene understanding)

• complexity of the objects
• environments changes
• configuration and relationships between objects

Prediction (anticipate future agent behaviors)
Lots of factors to consider:

• past behaviors
• scene semantic
• object’s curves
• other agents -> interaction

Planning (generate vechicle behaviors)

• safty first
• comfortable
• right signals to other particpants
• make progress on the route

3.2 Machine Learning Factory

Then he talked about Waymo’s maching learning factory. Which involving lots of Automatic Machine Learning algorithms.

In my point of view, there are three parts :

• Use automony in labeling data
• Auto tunes the model (architecture search, hyperparameters’ tunes)
• Transfer leaning stuff.

3.3 Simulation

I think this is the most important part of his speech. He elebrate they system of Simulation tests, why they doing it and how.

Why?
Even though we have a lot of data for driving tests. We still need for tests for some specially cases. Which have the least data, while having the greatest importance.

How?
They build a end-to-end deep learning based driver agent. And train on the specially cases with simulation in computer and also in real life scene.

Some keywords:

• Chauffeur net representation (a unit representation of all the factors in the scene)
• The DAGGER problem：if we only use the log data, and tell the system to imitate it. The system will be easy to make small errors overtime, and finally end up with a large distribution of probabilities (Chaos).
• Synthesizing pertubations
• Agent RNN with constraints.
• Long tail testing challenge.
• Smart agent build.

4. Technology Policy and Vehicle Safety

Given by Chris Gerdes from Stanford. He is the chef of US Department of Transportation. And part of the member to develop the federal automated vehicle policy.

bilibili video

4.1 Vehicle Safety Today

• This book is kind of the start of this standard. The author indicated that he was actually advocating things like airbags and anti-lock brakes. He thought the vehicles are lack of imagnation. If we have a safe system (or safe vehicle) we should not have such safe belts or other redendent staffs. (For example, the cars today are built extremely heavy to pass the crashing test. We can see further in the later speech that if our future automony vehicles are safe enough, we should reconsider these standards that we are following to build a car.)
• So we made this standard.
• The making of these rules are very time consuming. And it is really problematic, we are making rules for all the future cars, How could we descide how the future cars should be?
• For the current cars! are these car design legal?
• From FMVSS Scan, there are not much fro the ferdal level to prevent a automony vehicles from hiding roads.

4.2 Federal Automated Vehicles Policy

Not to tell the industy how to do sth, but rather suggest them that these are the things we would like you to address, and want the industry to come to the Federal departement to explain how they realize it.
-> to come with what is suitable for future car industry.

4.3 15 Point Safety Assessment

Operational Design Domain :
where the system operate, how to keep the system stays here, and what the callback if the car is outof the domain. How people interact with the car.

Validation Methods :
How to test the vehicles’ system.

4.4 Ethical considerations

Trolley car problem

• In the engineer point of view (to reduce risk). The car could choose a third solution to make a strong break. And if the break system is down, we could put some kind of sign

• Philosophical merites.

• Industry are actually adressing this problem (different break system, w.r.t whether the obstacle is car of human).

• Whether to keep the legal code or be more like human driver.
  
  some care stoped in front of you, should the car keep the code, and do not pass the double yellow lines, or more like human, go pass the lines to contiune to go forwards.

4.5 Data sharing