Even counter intuitive, Self-Driving service is actually a very human intensive operation. It’s been a key scaling problem for every Self-Driving startup. Waymo has a smart way to solve it. Anyone still dares to operate Robotaxi service, psst Tesla. Better watch & learn.
Eye in the Sky
Imagine you have the chance to stop two suicide bombers who are about to wreak havoc. However, there’s a catch: saving many lives means sacrificing an innocent girl. To pull or not to pull the trigger — that’s the dilemma. What would you decide? This is the gripping story of the movie Eye in the Sky. It’s a thought-provoking film worth watching, and it got me thinking 🤔.
Prediction, Judgement & Execution
In the age of AI, you may get all the intelligent you need. You may have the state of the art AI to predict the possible outcomes with a 99% confidence level. But, at the end of the day, someone still needs to make the judgment call. The decision is particularly tough when the stakes are high and time is running out.
Let’s say your judgment is nearly perfect, also at 99%. Then, someone else can execute the decision almost flawlessly, again at 99%. Now, you have a 97% (=99%*99%*99%) chance of “killing an innocent girl” to save many more lives from suicide bombings. Otherwise, you might miss the window of opportunity and lose many lives, likely children included. What do you do? 🤔
I can only hope someone is willing to take responsibility and make the right judgment call before it’s too late. I’m not comfortable delegating that to a machine just yet. We judge people on their decisions and actions all the time, and sometimes, we can hold them accountable when we’re lucky. But how can you “judge a machine”? 🤔
Solving Self-Driving’s “Trolley Problem
I’m pretty sure I’m a better-than-average driver. However, in my wife’s eyes, I’m not. She often reminds me not to get too close to the car in front or to steer back toward the middle of my lane 😅. Thanks to her training, I’ve started to believe “backseat drivers” might be part of the solution.
There is no 100% safe self-driving system yet. And I bet, there will never be one anytime soon. The question is how an self-driving system handles the 0.001% corner cases. Psst! In 2022, OpenAI shared how Reinforcement Learning from Human Feedback (RLHF) can make ChatGPT “smarter” to follow the instruction better. So, how may RLHF help the self-driving trolley problem?
Waymo: Call Judge Judy
In 2016, Google got a patent for remote assistance for autonomous vehicles in low-confidence situations. Key innovations included confidence-based triggers for assistance, multi-modal requests for assistance, and a GUI for remote operation. The remote assistance could come from various sources, including a human operator, a passenger within the vehicle, or even a more powerful remote AI system.
Fast forward to 2024, and Waymo has shared how its fleet response system is implemented. When Waymo Drivers (self-driving car technology) encounter ambiguous situations on the road, they can reach out to a human fleet response agent for additional information to unblock new paths forward.
Highlights of Waymo’s Fleet Response:
- Waymo Drivers and human fleet response agents primarily communicate through questions and answers.
- The fleet response tool provides real-time insights into a 3D scene where a Waymo vehicle is operating. The human agent can add new route guidance to help the Waymo Driver navigate a new path.
- The human agent can assist the Waymo Driver by providing additional context as it navigates scenarios involving active emergency vehicles.
Advantages of Waymo’s Fleet Response:
- Learning from Human Creativity: Waymo Drivers can learn from human agents’ creative solutions, making all Waymo Drivers smarter over time.
- Significant Cost Reduction: The system significantly reduces operational costs because each human agent can support multiple self-driving vehicles simultaneously. In contrast, Robotaxi operators that rely on human drivers to take over remotely cannot scale up efficiently and struggle to compete.
