Show HN: Flyingcar.pet – a physics simulation of Tesla's flying Roadster stunt
flyingcar.pet
I gave Opus one task: look at the video and create a simulation like that using three.js, matching the video as closely as possible. I'd thought that Opus would one- or two-shotted this since in my mind it was just estimating height and distances and working backwards to compute the thrusts. Little did I know this would turn into a multi-day project.
Opus's first hurdle was recognizing the five parked Tesla vehicles. The video was dark, so it was hard for Opus to see. I helped it identify the models and also gave it the approximate dimensions of all six Tesla vehicles, including the Roadster.
1. The initial version looked close enough to the video, but Opus made the car fly straight, whereas the video shows the Roadster veering to the left and landing to its left, so I asked Opus to watch the video again and correct that.
2. The next version showed the car drifting to the left mid-flight like the video, but the position was off when I compared frame by frame to the video. It turned out that Opus was being lazy and used only a few frames as reference instead of doing it frame by frame.
3. Once Opus matched the car position to the video, frame by frame, I realized that the roll did not match, so I asked Opus to investigate and redo.
4. The next version had the position and roll correct, but I did not see any thrusters firing during flight. It turned out that Opus was creating a smooth trajectory where the thrusters only fired for the launch and the landing, and stayed off mid-flight. This was not what I'd seen in the video, however.
5. Opus then changed it so that the thrusters would fire during the flight, which would affect the trajectory continuously. However, at this point I realized something odd: it was firing a thruster upward, which is in the opposite direction of the four bottom-mounted thrusters. It was definitely strange since the video did not show any upward-firing thrusters. It turned out that Opus was using 12 thrusters in total because it relied on Elon Musk's comment in 2018 that the Roadster would have around 10 thrusters. Using 12 thrusters was what allowed it to create any trajectory it wanted to. I asked it to use only 6 thrusters: four bottom ones for lifting and landing, and two side thrusters mounted near the front wheels since those were the ones I could see firing in the video.
6. Opus gave me another simulation with 6 thrusters firing, but now I could not see the side thrusters firing at all during the flight, which is similar to #4. It turned out that up to this point, Opus had been doing the simulation under the ideal conditions where the trajectory was pre-computed instead of doing a closed-loop control and constantly adjusting the thrust to keep the car on the planned path. Opus then added wind gusts and redid the simulation using a 100 Hz feedback loop.
At this point, I have something that is close enough to the video. It is still not perfect since I would love to match the firing sequence, intensity, and cloud shape of the thrusters during the flight, but I think this is a good stopping point for me, and now I have something where I can tweak the params like the car's weight, max thruster force, max height, number of rolls, etc. and model any flight path I'd like. There are a lot of params to configure, and some combinations can result in invalid flight paths, so I added WebMCP so that we can use the Claude plugin for Chrome to drive the simulation and ask the AI any questions. I found that Sonnet 5.5 Medium is good enough for this. Haiku 5.5 Medium is fast but also makes mistakes.
Caveats: the mass, drag coefficients, thruster positions and gas performance are estimates, not Tesla data. The closed-loop model also assumes perfect sensors.