DiffIDM [Arxiv]
This is the official code of our paper "Gradient-based Trajectory Optimization with Parallelized Traffic Simulation", which has been accepted to ICRA 2025. Here we provide the code for parallelized differentiable IDM computation layer, which could be easily adopted for larger scale traffic simulators. The computation layer can handle upto 2 million vehicles in real time using either CPU or GPU. We also provide our experiment code for filtering vehicle trajectories in NGSim dataset using our simulator. Please see our paper for more details.
You need to install pytorch to use our computation layer. Then, you can install our computation layer using pip.
pip install diffidmAfter installation, you can use the computation layer as follows.
import torch
from diffidm import IDMLayer
device = 'cuda' if torch.cuda.is_available() else 'cpu'
num_vehicles = int(1e6) # 1M vehicles
print("Device:", device)
print("Number of vehicles:", num_vehicles)
### randomly generate IDM variables
a_max = torch.rand(num_vehicles, device=device) * 5 + 5 # [5, 10], maximum acceleration
a_min = torch.rand(num_vehicles, device=device) * 5 - 10 # [-10, -5], minimum acceleration
a_pref = torch.rand(num_vehicles, device=device) * 4.9 + 0.1 # [0.1, 5], preferred acceleration
v_curr = torch.rand(num_vehicles, device=device) * 40 + 20 # [20, 60], current velocity
v_target = torch.rand(num_vehicles, device=device) * 40 + 20 # [20, 60], target velocity
pos_delta = torch.rand(num_vehicles, device=device) * 10 + 5 # [5, 15], headway distance to the leading vehicle
vel_delta = torch.rand(num_vehicles, device=device) * 20 + 10 # [10, 30], relative velocity to the leading vehicle
min_space = torch.rand(num_vehicles, device=device) * 9 + 1 # [1, 10], minimum space headway
time_pref = torch.rand(num_vehicles, device=device) * 4.9 + 0.1 # [0.1, 5], desired time headway
delta_time = torch.full((num_vehicles,), 0.01, device=device) # 0.01, simulation time step
a_max = a_max.requires_grad_()
a_min = a_min.requires_grad_()
a_pref = a_pref.requires_grad_()
v_curr = v_curr.requires_grad_()
v_target = v_target.requires_grad_()
pos_delta = pos_delta.requires_grad_()
vel_delta = vel_delta.requires_grad_()
min_space = min_space.requires_grad_()
time_pref = time_pref.requires_grad_()
delta_time = delta_time.requires_grad_()
print("IDM variables generated.")
### forward pass: compute acceleration using IDM
start_event = torch.cuda.Event(enable_timing=True)
end_event = torch.cuda.Event(enable_timing=True)
start_event.record()
acc = IDMLayer.apply(
a_max,
a_min,
a_pref,
v_curr,
v_target,
pos_delta,
vel_delta,
min_space,
time_pref,
delta_time,
)
end_event.record()
torch.cuda.synchronize()
elapsed_time = start_event.elapsed_time(end_event)
print(f"Forward pass time: {elapsed_time} ms")
### backward pass: compute gradients of IDM variables
start_event.record()
acc.sum().backward()
end_event.record()
torch.cuda.synchronize()
elapsed_time = start_event.elapsed_time(end_event)
print(f"Backward pass time: {elapsed_time} ms")If installed correctly, it would print as follows.
Device: cuda
Number of vehicles: 1000000
IDM variables generated.
Forward pass time: 1.222208023071289 ms
Backward pass time: 3.7359039783477783 msNow you can use it in your code!
As demonstrated in our paper, we can use our computation layer to filter physically unrealistic vehicle motions from their trajectories captured in real-world. We tested our filtering algorithm on NGSIM dataset. To run the algorithm, first download the dataset (1.53GB) from this page to data directory, under the name of ngsim.csv. Then, preprocess the data with the following command.
python ngsim_preprocess.pyThis preprocessing step generates input data for our filtering algorithm, which would be saved in input/. Then, we can use the following command to filter the NGSIM trajectories in us-101 category using our computation layer. Please see ngsim_filter.py for the details abour arguments.
CUDA_VISIBLE_DEVICES=0 OMP_NUM_THREADS=1 python ngsim_filter.py --token=us-101 --device=cuda The filtered trajectories and related optimization results (e.g. IDM parameters for each trip) are stored under output/ directory. Finally, we can evaluate the filtered trajectories by comparing them with the ground truth data and generate the experimental results in the paper with following command. Also see ngsim_eval.py for the details about arguments.
python ngsim_eval.py --token=us-101 --render-traj-id=1000The evaluation results are saved in eval/.
If you found our work to be useful, please consider citing our work.
@article{son2024gradient,
title={Gradient-based Trajectory Optimization with Parallelized Differentiable Traffic Simulation},
author={Son, Sanghyun and Zheng, Laura and Clipp, Brian and Greenwell, Connor and Philip, Sujin and Lin, Ming C},
journal={arXiv preprint arXiv:2412.16750},
year={2024}
}We used Wei Ma's NGSIM interface code for preprocessing NGSIM dataset. We appreciate this great work.