Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Faithful Dynamic Imitation Learning from Human Intervention with Dynamic Regret Minimization
Authors: Bo Ling, Zhengyu Gan, Wanyuan Wang, Guanyu Gao, Weiwei Wu, Yan Lyu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on Meta Drive and CARLA driving benchmarks demonstrate that Faith Da IL achieves state-of-the-art performance in safety and task success with significantly reduced human intervention data compared to prior HIL baselines. We conducted extensive experiments on Meta Drive and CARLA driving benchmarks. Results show Faith Da IL significantly outperforms leading HIL baselines in safety and task success, with notably less human intervention data. We conducted ablation studies to assess effectiveness of Faith Da IL s key components: Dynamic Regret Minimization (DRM) and the Faithful Off-policy imitation learning module (FOP). |
| Researcher Affiliation | Academia | Bo Ling Southeast University EMAIL Zhengyu Gan Southeast University EMAIL Wanyuan Wang Southeast University EMAIL Guanyu Gao Nanjing University of Science and Technology EMAIL Weiwei Wu Southeast University EMAIL Yan Lyu Southeast University EMAIL |
| Pseudocode | Yes | Algorithm 1 Faithful Dynamic Imitation Learning from Human Intervention (Faith Da IL) |
| Open Source Code | Yes | The corresponding source code is available at https://github.com/William-island/Faith Da IL. We release the full implementation of our algorithm in the supplementary material, and the corresponding source code is available with anonymized URL of https://anonymous.4open.science/r/nips2025-DC57. |
| Open Datasets | Yes | Environments. We evaluate our approach on two challenging driving simulators: Meta Drive Safety Benchmark [46] and CARLA Town01 [47]. |
| Dataset Splits | Yes | We created training and testing sets of 100 distinct scenarios to assess generalization. Training scenarios were randomly selected from 100 diverse environments in Meta Drive and 25 varied routes with different start/end points, lighting, and weather conditions in CARLA Town01. |
| Hardware Specification | Yes | Experiments were run on a machine with an Nvidia Ge Force RTX 3070 Ti Laptop GPU and an Intel Core i7-12700H CPU, supporting real-time simulation and training. |
| Software Dependencies | No | Our implementation builds upon open-source repositories of ODICE [25], PVP [9] and FTPL-D+ [40]. Implementations of PPO and TD3 utilize Stable-Baselines3 [48], while other HIL baselines use official implementations where available [9, 11]. No specific version numbers for software dependencies are provided in the main text. |
| Experiment Setup | No | Hyper-parameter and other details are in Appendix D. The main text (Section 5.1) only states that hyperparameters are in Appendix D, without providing the specific values. |