Leaderboard · comma.ai/jobs · Discord · X
Machine learning models can drive cars, paint beautiful pictures and write passable rap. But they famously suck at doing low level controls. Your goal is to write a good controller. This repo contains a model that simulates the lateral movement of a car, given steering commands. The goal is to drive this "car" well for a given desired trajectory.
This neural network based controller outperforms all other known controllers. This graph illustrates the cost of all 5000 simulations across a variety of controllers, to compare their performance.
Instead of relying only on PID data for training this model, I built a 2D racing game to collect human driving data. This is compatible with the keyboard and a Logitech steering wheel. This allows me to collect more realistic human driving data, for the most complex and dynamic trajectories.
- Improved graphing to include future states as dotted lines. This was useful to view the stability of future states, before deciding to build a 2D racing sim based mostly on these future values.
- Improved the simulator to pass the initial steer values to each controller, to help controllers smoothly take over control of the vehicle.
- Improved random number generation to be multi-thread friendly, so we can run multiple simulations at the same time.
This was not a very straight-forward process, lol! Finding the "right" distribution of training files, model architecture, and optimization was a real learning process. I had many breakthroughs along the way:
- The right input parameters (15 to be exact: 3 lataccel, 3 roll, 3 v_ego, 3 a_ego, and 3 previous steer torques)
- The right output parameters (the next 2 steer torques)
- The right window size (30 input states - 10 per second - 3 seconds of input data)
- For more info on training the neural network.
Proof that a neural network can surpass all other PID controller designs. Read the full results here.
We'll be using a synthetic dataset based on the comma-steering-control dataset for this challenge. These are actual car and road states from openpilot users.
# install required packages
# recommended python==3.11
pip install -r requirements.txt
# test this works
python tinyphysics.py --model_path ./models/tinyphysics.onnx --data_path ./data/00000.csv --debug --controller pid
There are some other scripts to help you get aggregate metrics:
# batch Metrics of a controller on lots of routes
python tinyphysics.py --model_path ./models/tinyphysics.onnx --data_path ./data --num_segs 100 --controller pid
# generate a report comparing two controllers
python eval.py --model_path ./models/tinyphysics.onnx --data_path ./data --num_segs 100 --test_controller pid --baseline_controller zero
You can also use the notebook at experiment.ipynb for exploration.
This is a "simulated car" that has been trained to mimic a very simple physics model (bicycle model) based simulator, given realistic driving noise. It is an autoregressive model similar to ML Controls Sim in architecture. Its inputs are the car velocity (v_ego), forward acceleration (a_ego), lateral acceleration due to road roll (road_lataccel), current car lateral acceleration (current_lataccel), and a steer input (steer_action), then it predicts the resultant lateral acceleration of the car.
Your controller should implement a new controller. This controller can be passed as an arg to run in-loop in the simulator to autoregressively predict the car's response.
Each rollout will result in 2 costs:
-
lataccel_cost:$\dfrac{\Sigma(actual\_lat\_accel - target\_lat\_accel)^2}{steps} * 100$ -
jerk_cost:$\dfrac{\Sigma((actual\_lat\_accel_t - actual\_lat\_accel_{t-1}) / \Delta t)^2}{steps - 1} * 100$
It is important to minimize both costs. total_cost:
Run the following command, then submit report.html and your code to this form.
python eval.py --model_path ./models/tinyphysics.onnx --data_path ./data --num_segs 5000 --test_controller <insert your controller name> --baseline_controller pid
- With this commit we made the simulator more robust to outlier actions and changed the cost landscape to incentivize more aggressive and interesting solutions.
- With this commit we fixed a bug that caused the simulator model to be initialized wrong.
Like this sort of stuff? You might want to work at comma! comma.ai/jobs