TITRATED: Learned Human Driving Behavior without Infractions via Amortized Inference
Authors: Vasileios Lioutas, Adam Scibior, Frank Wood
TMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate this titration technique using the ITRA foundation behavior model trained on the INTERACTION dataset when transferring to CARLA map locations. We demonstrate a 76-86% reduction in infraction rate and provide evidence that further gains are possible with more computation or better inference algorithms. In our experiments, we, at a high level, learn human driving behaviors from the INTERACTION dataset (Zhan et al., 2019) and use them to create NPCs in CARLA (Dosovitskiy et al., 2017). Table 2 shows that TITRATED is able to significantly reduce the collision and off-road rates of ITRA. |
| Researcher Affiliation | Collaboration | Vasileios Lioutas EMAIL Inverted AI University of British Columbia Adam Ścibior EMAIL Inverted AI Frank Wood EMAIL Inverted AI University of British Columbia Montréal Institute for Learning Algorithms (Mila) |
| Pseudocode | Yes | Algorithm 1 Infraction-Free Dataset Generation Algorithm 2 Amortized Inference with Infraction Penalties Training Process |
| Open Source Code | No | The paper does not explicitly state that the code is open-source, nor does it provide a link to a code repository. It mentions "TITRATED is already used to provide NPCs for CARLA based on human driving trajectories recorded in the real world." but this is not a statement of code release for the methodology. |
| Open Datasets | Yes | In this paper, we use ITRA (Ścibior et al., 2021) trained on the INTERACTION dataset (Zhan et al., 2019) as our foundation model (Bommasani et al., 2021) and target the creation of NPCs for CARLA (Dosovitskiy et al., 2017). |
| Dataset Splits | Yes | To create a collection of initial conditions required by Algorithm 1 in CARLA, we created a custom synthetic dataset using the built-in autopilot. Specifically, in each town we recorded 200 minutes of driving at 10 frames per second, featuring 100 randomly selected and placed vehicles. We then selected 23 locations of interest across 7 towns and cropped the recordings to a 100 meter radius from each of them, mimicking the structure of the INTERACTION dataset. We then sliced those datasets into 4 second segments, consisting of 1 second of history and 3 seconds of predictions, taking the 1 second history from those segments as initial conditions. The dataset is summarized in Table 1. Table 1: Summary of the dataset collected using CARLA autopilot. These are the initial conditions used to adapt TITRATED to CARLA. Scene Train / Val. Size Town01_Straight 7338 / 492 |
| Hardware Specification | Yes | The batch size is set to 8 trajectories using class-mates forcing (Ścibior et al., 2021; Tang & Salakhutdinov, 2019) and trained on a single NVIDIA RTX 2080 Ti GPU. |
| Software Dependencies | No | The paper mentions using the ADAM optimizer and refers to ITRA and CARLA, but it does not specify version numbers for any programming languages, libraries, or other software dependencies crucial for replication (e.g., PyTorch version, Python version). |
| Experiment Setup | Yes | We follow the same hyper-parameter setup as described in Ścibior et al. (2021) and fine-tune a mirror copy of the ITRA model for each unseen location from the synthetic dataset using the proposed methodology. Specifically, we set the model hidden size dh to 64, the latent representation size dz to 2 and the number of recurrent layers to 2. We use differentiable rendering as described in Ścibior et al. (2021) and a birdview image of size 256 256. We use the ADAM optimizer (Kingma & Ba, 2015) with default values and 1 10 3 learning rate. The batch size is set to 8 trajectories using class-mates forcing (Ścibior et al., 2021; Tang & Salakhutdinov, 2019) and trained on a single NVIDIA RTX 2080 Ti GPU. We set the hyperparameters λC and λOR to 1 103 and 1 102 respectively and use max_trials of 10 and just-in-time sampling of infraction-free training examples. |