Long-Horizon Visual Planning with Goal-Conditioned Hierarchical Predictors

Authors: Karl Pertsch, Oleh Rybkin, Frederik Ebert, Shenghao Zhou, Dinesh Jayaraman, Chelsea Finn, Sergey Levine

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The aim of our experiments is to study the following questions: (1) Are the proposed GCPs able to effectively predict goal-directed trajectories in the image space and scale to long time horizons? (2) Is the proposed goal-conditioned hierarchical planning method able to solve long-horizon visual control tasks? (3) Does the version of GCP with adaptive binding find high-level events in the trajectories?
Researcher Affiliation Academia Karl Pertsch ,1 Oleh Rybkin ,2 Frederik Ebert3 Chelsea Finn4 Dinesh Jayaraman2 Sergey Levine3 1 USC 2 UPenn 3 UC Berkeley 4 Stanford University
Pseudocode Yes Algorithm 1 Goal-Conditioned Hierarchical Planning
Open Source Code Yes Project page: orybkin.github.io/video-gcp
Open Datasets Yes We further evaluate on the real-world Human 3.6M video dataset [25]
Dataset Splits No The paper discusses training data and evaluation, but does not explicitly specify exact percentages or sample counts for training, validation, and test splits. It mentions collecting 'example trajectories' but does not detail the splits.
Hardware Specification Yes We use 16 16 px to fit the 1000-frame sequences on a single NVIDIA V100 GPU
Software Dependencies No The paper refers to various methods and models, but does not provide specific version numbers for software dependencies or libraries used in their implementation (e.g., PyTorch, TensorFlow versions, or specific optimizer versions beyond Adam).
Experiment Setup No The paper mentions architectural details and hyperparameter considerations but defers specific details to the appendix: 'Architecture and hyperparameters are detailed in Appendix C.' Since the question asks for details *in the main text*, the answer is No for specific values.