Learning to Discover Skills through Guidance

Authors: HYUNSEUNG KIM, BYUNG KUN LEE, Hojoon Lee, Dongyoon Hwang, Sejik Park, Kyushik Min, Jaegul Choo

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical evaluation demonstrates that DISCO-DANCE outperforms other USD baselines in challenging environments, including two navigation benchmarks and a continuous control benchmark.
Researcher Affiliation Collaboration 1Kim Jaechul Graduate School of AI, KAIST. 2KAKAO Corp.
Pseudocode Yes Algorithm 1: Skill Discovery through Guidance
Open Source Code Yes Qualitative visualizations and code of DISCO-DANCE are available at https://mynsng.github.io/discodance/.
Open Datasets Yes 2D mazes: We used the publicly available environment code in [7]. Ant mazes: We used the publicly available environment code in [8]. Deepmind Control suite: We used the publicly available environment code in [22, 53].
Dataset Splits No The paper describes training steps and finetuning steps in RL environments, but does not specify distinct training, validation, and test *dataset splits* with percentages or sample counts.
Hardware Specification Yes We conducted all experiments using a single RTX 3090 GPU, and each experiment requires approximately up to 2GB of memory.
Software Dependencies No The paper mentions software like SAC, DIAYN, APS, SMM, and URLB, and specific environment code, but does not provide specific version numbers for these software components or programming languages.
Experiment Setup Yes Table 4: A common set of hyperparameters used in DISCO-DANCE. Table 5: Per environment sets of hyperparameters used in DISCO-DANCE.