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..

Coarse-to-fine Q-Network with Action Sequence for Data-Efficient Reinforcement Learning

Authors: Younggyo Seo, Pieter Abbeel

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments show that CQN-AS outperforms several baselines on a variety of sparse-reward humanoid control and tabletop manipulation tasks from Bi Gym and RLBench.
Researcher Affiliation Academia Younggyo Seo UC Berkeley EMAIL Pieter Abbeel UC Berkeley EMAIL
Pseudocode No The paper describes the algorithms and methods in textual form within sections like 'Coarse-to-fine Q-Network with Action Sequence' and 'Full description of CQN and CQN-AS', along with architectural diagrams, but does not present a dedicated pseudocode or algorithm block.
Open Source Code Yes Code is available at: https://younggyo.me/cqn-as/
Open Datasets Yes Our experiments show that CQN-AS outperforms several baselines on a variety of sparse-reward humanoid control and tabletop manipulation tasks from Bi Gym and RLBench. ... Bi Gym (Chernyadev et al., 2024) and RLBench (James et al., 2020). ... Humanoid Bench (Sferrazza et al., 2024).
Dataset Splits Yes To focus on challenging robotic tasks that aim to induce policies generating realistic behaviors, we consider a practical setup of demo-driven RL where we initialize training with a modest amount of expert demonstrations and then train with online data. ... We consider 25 Bi Gym tasks with 17 to 60 demonstrations. ... We use 100 demonstrations generated via motion-planning.
Hardware Specification Yes For Bi Gym and Humanoid experiments, we use NVIDIA A5000 GPU with 24GB VRAM. ... For RLBench experiments, we use NVIDIA RTX 2080Ti GPU, with which each experiment with 30K environment steps take 6.5 hours.
Software Dependencies No The paper describes various architectural components (e.g., CNN, MLP, GRU, SiLU, Layer Norm) and optimizers (AdamW) along with their corresponding research papers, but does not provide specific version numbers for software libraries or frameworks (e.g., Python, PyTorch, CUDA) used for implementation.
Experiment Setup Yes We use the same set of hyperparameters across the tasks in each domain. For hyperparameters shared across CQN and CQN-AS, we use the same hyperparameters for both algorithms for a fair comparison. We provide detailed hyperparameters for Bi Gym and RLBench experiments in Table 3 and Humanoid Bench experiments in Table 4.