GenPlan: Generative Sequence Models as Adaptive Planners

Authors: Akash Karthikeyan, Yash Vardhan Pant

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the effectiveness of our method through multiple simulation environments. Notably, Gen Plan outperforms state-of-the-art methods by over 10% on adaptive planning tasks, where the agent adapts to multitask missions while leveraging demonstrations from single-goal-reaching tasks. ... We evaluate the performance of Gen Plan in Baby AI (Chevalier-Boisvert et al. 2019) and continuous manipulation tasks, focusing on the agent s adaptive and generalization capabilities. ... We conduct simulations in a modified Baby AI suite following three paradigms.
Researcher Affiliation Academia University of Waterloo, Canada EMAIL
Pseudocode Yes Algorithm 1: Gen Plan Training ... Algorithm 2: Gen Plan Sampling
Open Source Code Yes Code https://github.com/CL2-UWaterloo/Gen Plan
Open Datasets Yes We evaluate the performance of Gen Plan in Baby AI (Chevalier-Boisvert et al. 2019) and continuous manipulation tasks: (a) Push T (Florence et al. 2021) and (b) Franka Kitchen (Gupta et al. 2019).
Dataset Splits No The model, trained on simple goal-reaching tasks, is evaluated for zero-shot adaptation across harder environments without additional finetuning. ... Success rates in reaching goals and completing tasks are reported across 250 novel environments. The map layout, goals, obstacles, and agent positions are randomized in each run.
Hardware Specification Yes We implemented Gen Plan using Python 3.8 and trained it on a 12-core CPU alongside an RTX A6000 GPU.
Software Dependencies No The paper mentions 'Python 3.8' but does not list multiple key software components with their versions or a self-contained solver/specialized package with a specific version number.
Experiment Setup Yes The context length is typically set to 1 but can be extended to increase the agent s memory (see figure 2B). ... The entropy lower bound β in eq. 4b is currently a hyper-parameter that must be manually specified. ... For model and environment hyperparameters, we adopt the configurations from (Lee et al. 2024).