Online Learning and Control of Complex Dynamical Systems from Sensory Input

Authors: Oumayma Bounou, Jean Ponce, Justin Carpentier

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The proposed approach is evaluated empirically on several classical dynamical systems and sensory modalities, with good performance on long-term prediction and control.
Researcher Affiliation Academia 1Inria and Département d Informatique de l Ecole Normale Supérieure, PSL Research University 2Center for Data Science, New York University
Pseudocode No The paper does not contain explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes 3. (a) Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)? [Yes] see supplemental material
Open Datasets No We generate three black-and-white 64 64 video datasets of classical dynamical systems using the Pinocchio library [49] for the simulation and Panda3D-viewer [50] for the rendering: a simple pendulum, a double pendulum1 and a cartpole. ... More details on the datasets generation can be found in the supplementary material.
Dataset Splits No The paper mentions "training datasets" and evaluation on sequences of different lengths (e.g., "trained on 5 s sequences" and "evaluated on 15 s sequences") but does not explicitly specify train/validation/test splits or mention a "validation" set.
Hardware Specification Yes Each learning problem takes about 3 hours to solve on an Nvidia RTX6000 GPU.
Software Dependencies No The paper mentions software like "Pinocchio library [49]", "Panda3D-viewer [50]", and "Py Torch" but does not specify their version numbers.
Experiment Setup Yes The dimension of the codes in latent space corresponds to the bottleneck size of the autoencoder, set to nz = 8 in all our experiments. More details on the architecture can be found in the supplementary material. The models of the unactuated systems are trained on 5 s sequences and evaluated on 15 s sequences, and those of actuated systems were trained on 10 s sequences and evaluated on 20 s sequences.