Compositional Plan Vectors

Authors: Coline Devin, Daniel Geng, Pieter Abbeel, Trevor Darrell, Sergey Levine

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments aim to understand how well CPVs can learn tasks of varying complexity, how well they can generalize to tasks that are more complex than those seen during training (thus demonstrating compositionality), and how well they can handle additive composition of tasks, where the policy is expected to perform both of the tasks in sequence. We hypothesize that, by conditioning a policy on the subtraction of the current progress from the goal task embedding, we will learn a task representation that encodes tasks as the sum of their component subtasks.
Researcher Affiliation Academia University of California, Berkeley
Pseudocode No The paper describes network architectures and processes, but does not contain explicitly labeled pseudocode or algorithm blocks.
Open Source Code No Along with the environments, we will release code to generate demonstrations of the compositional tasks.
Open Datasets No We introduce two new learning environments, shown in Figures 3 and 4, that test an agent s ability to perform tasks that require different sequences and different numbers of sub-skills. Along with the environments, we will release code to generate demonstrations of the compositional tasks.
Dataset Splits Yes The data is divided into training and validation sets 90/10.
Hardware Specification Yes All models are trained on either k-80 GPUs or Titan X GPUs.
Software Dependencies No The paper mentions implicit software usage (e.g., neural networks), but does not provide specific software dependencies with version numbers.
Experiment Setup Yes We compared all models across embedding dimension sizes of [64,128,256, and 512]. In the crafting environment, the 512 size was best for all methods. In the grasping environment, the 64 size was best for all methods. For TECNets, we tested λctr = 1 and 0.1, and found that 0.1 was best.