SkiLD: Unsupervised Skill Discovery Guided by Factor Interactions

Authors: Zizhao Wang, Jiaheng Hu, Caleb Chuck, Stephen Chen, Roberto Martín-Martín, Amy Zhang, Scott Niekum, Peter Stone

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate Ski LD in several domains with challenging, long-horizon sparse reward tasks including a realistic simulated household robot domain, where Ski LD successfully learns skills with clear semantic meaning and shows superior performance compared to existing unsupervised reinforcement learning methods that only maximize state coverage.
Researcher Affiliation Collaboration 1University of Texas at Austin 2University of Massachusettes, Amherst 3Sony AI
Pseudocode Yes Algorithm 1 Ski LD Skill Discovery
Open Source Code Yes Code and visualizations are at https://wangzizhao.github.io/Ski LD/.
Open Datasets Yes We evaluate our method on two challenging object-rich embodied AI benchmarks: Mini-behavior [32] and Interactive Gibson [42].
Dataset Splits No The paper describes pre-training and finetuning steps, but does not explicitly mention a 'validation' dataset split in the context of fixed datasets or provide specific details for such a split beyond training and evaluation phases.
Hardware Specification Yes Nvidia A40 GPU; Intel(R) Xeon(R) Gold 6342 CPU @2.80GHz Nvidia A100 GPU; Intel(R) Xeon(R) Gold 6342 CPU @2.80GHz
Software Dependencies No The codebase is built on tianshou [64] for backend RL, though with significant modifications.
Experiment Setup Yes The hyperparameters of skill learning and task learning can be found in Table 1.