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..
Interactive and Hybrid Imitation Learning: Provably Beating Behavior Cloning
Authors: Yichen Li, Chicheng Zhang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct experiments in Mu Jo Co continuous control tasks and show that if the cost of state-wise interactive demonstration is not much higher than its offline counterpart, interactive algorithms can enjoy a better cost efficiency than Behavior Cloning. Under some cost regimes and some environments, hybrid imitation learning can outperform approaches that use either source alone. |
| Researcher Affiliation | Academia | Yichen Li University of Arizona EMAIL Chicheng Zhang University of Arizona EMAIL |
| Pseudocode | Yes | Algorithm 1 STAGGER: DAgger with State-wise annotation oracle Algorithm 2 WARM-STAGGER: Warm-start STAGGER with offline demonstrations Algorithm 3 TRAGGER: DAgger with trajectory-wise annotation oracle Algorithm 4 WARM-TRAGGER: Warm-start TRAGGER with offline demonstrations |
| Open Source Code | Yes | Our implementation builds on the publicly available DRIL repository [7] (https://github.com/xkianteb/dril), with modifications to support interactive learning. The continuous control environments used in our experiments are: Half Cheetah Bullet Envv0 , Ant Bullet Env-v0 , Walker2DBullet Env-v0 , and Hopper Bullet Env-v0 . We include link to our implementation here: https://github.com/liyichen1998/Interactive-and-Hybrid-Imitation-Learning-Provably-Beating-Behavior-Cloning. |
| Open Datasets | Yes | We conduct a simple simulation study comparing the sample efficiency of log-loss Behavior Cloning [17] and STAGGER in four Mu Jo Co [72, 8] continuous control tasks with H = 1000 and pretrained deterministic MLP experts [52, 53]. |
| Dataset Splits | Yes | Each model is trained from random initialization using a batch size of 100, a learning rate of 10 3, and up to 2000 passes over the dataset, with early stopping evaluated every 250 passes using a 20% held-out validation set. |
| Hardware Specification | Yes | All experiments were conducted on a Linux workstation equipped with an Intel Core i9 CPU (3.3GHz) and four NVIDIA Ge Force RTX 2080 Ti GPUs. |
| Software Dependencies | Yes | The continuous control environments used in our experiments are: Half Cheetah Bullet Envv0 , Ant Bullet Env-v0 , Walker2DBullet Env-v0 , and Hopper Bullet Env-v0 . |
| Experiment Setup | Yes | Each model is trained from random initialization using a batch size of 100, a learning rate of 10 3, and up to 2000 passes over the dataset, with early stopping evaluated every 250 passes using a 20% held-out validation set. |