Discrete Latent Plans via Semantic Skill Abstractions

Authors: Haobin Jiang, Wang, Zongqing Lu

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through experiments in simulated control environments, we demonstrate that LADS outperforms state-of-the-art methods in both skill learning and compositional generalization. The code is available at https://github.com/PKU-RL/LADS. 5 EXPERIMENTS
Researcher Affiliation Academia Haobin Jiang1, Jiangxing Wang1, Zongqing Lu1,2 1School of Computer Science, Peking University 2Beijing Academy of Artificial Intelligence Correspondence to Zongqing Lu <EMAIL>.
Pseudocode Yes We provide the pseudocode of LADS as shown in Algorithm 1. Algorithm 1 Training LADS
Open Source Code Yes The code is available at https://github.com/PKU-RL/LADS.
Open Datasets Yes LORe L (Nair et al., 2022) is a simulated domain developed on top of Meta-World (Yu et al., 2020)... Kitchen (Gupta et al., 2019) is a simulated domain developed on top of Mu Jo Co (Todorov et al., 2012)...
Dataset Splits Yes For testing, we selected 3 instructions, resulting in a training dataset of 22 instructions and 509 demonstrations.
Hardware Specification No No specific hardware details (like GPU models, CPU types, or memory amounts) used for running experiments are provided in the paper.
Software Dependencies No We use a pretrained Distil BERT (Sanh, 2019) as the language encoder and a causal transformer (Chen et al., 2021a) as the high-level policy. ... We use CLIP Vi T-B/32 to encode the instruction... For Kitchen (image), we use a pretrained Res Net18 (He et al., 2016) to encode the images...
Experiment Setup Yes The hyperparameters for the network architecture not covered in Appendix B.1, as well as those related to training, are listed in Table 4. Table 4: Hyperparameters of our experiments.