Deep Imitative Models for Flexible Inference, Planning, and Control
Authors: Nicholas Rhinehart, Rowan McAllister, Sergey Levine
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method using the CARLA driving simulator (Dosovitskiy et al., 2017). We seek to answer four primary questions: (1) Can we generate interpretable, expert-like plans with offline learning and minimal reward engineering? Neither IL nor MBRL can do so. It is straightforward to interpret the trajectories by visualizing them on the ground plane; we thus seek to validate whether these plans are expert-like by equating expert-like behavior with high performance on the CARLA benchmark. (2) Can we achieve state-of-the-art CARLA performance using resources commonly available in real autonomous vehicle settings? ... (3) How flexible is our approach to new tasks? ... (4) How robust is our approach to error in the provided goals? ... Results: Towards questions (1) and (3) (expert-like plans and flexibility), we apply our approach with a variety of goal likelihoods to the CARLA simulator. Towards question (2), we compare our methods against CILS, MBRL, and prior work. These results are shown in Table 3. |
| Researcher Affiliation | Academia | Nicholas Rhinehart UC Berkeley nrhinehart@berkeley.edu Rowan Mc Allister UC Berkeley rmcallister@berkeley.edu Sergey Levine UC Berkeley svlevine@berkeley.edu |
| Pseudocode | Yes | Pseudocode of the driving and inference algorithms are given in Algs 1 and 2. The PID algorithm is given in Appendix A. ... Algorithm 1 IMITATIVEDRIVING(...) ... Algorithm 2 IMITATIVEPLAN(...) ... Algorithm 3 PIDCONTROLLER(...) |
| Open Source Code | No | Videos are available at https://sites.google.com/view/imitative-models. ... We have prepared the dataset of collected data for public release upon publication. |
| Open Datasets | No | We have prepared the dataset of collected data for public release upon publication. |
| Dataset Splits | Yes | We randomized episodes to either train, validation, or test sets. We created sets of 60,701 train, 7586 validation, and 7567 test scenes, each with 2 seconds of past and 4 seconds of future position information at 10Hz. |
| Hardware Specification | No | The paper does not specify the hardware (e.g., GPU model, CPU model, RAM) used for running the experiments. |
| Software Dependencies | No | The paper mentions CARLA and R2P2, but does not provide specific version numbers for these or any other software dependencies required for replication. |
| Experiment Setup | Yes | We used T =40 trajectories at 10Hz (4 seconds), and τ =3. ... We started with N = 120 different z initializations, optimized them in batch, and returned the highest-scoring value across the entire optimization. ... ran the optimization for a small number of steps, M = 10, and found that we obtained good performance. ... We ran the autopilot in Town01 for over 900 episodes of 100 seconds each in the presence of 100 other vehicles, and recorded the trajectory of every vehicle and the autopilot s LIDAR observation. |