Latent Variable Sequential Set Transformers for Joint Multi-Agent Motion Prediction

Authors: Roger Girgis, Florian Golemo, Felipe Codevilla, Martin Weiss, Jim Aldon D'Souza, Samira Ebrahimi Kahou, Felix Heide, Christopher Pal

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our method with strong empirical results on diverse tasks, including the large-scale nu Scenes and Argoverse datasets for autonomous driving (Caesar et al., 2020; Chang et al., 2019), the multi-lingual Omniglot dataset of handwritten characters (Lake et al., 2015), and the synthetic partition of the Traj Net++ pedestrian trajectory prediction task (Sadeghian et al., 2018).
Researcher Affiliation Collaboration 1Polytechnique Montréal, 2Mila, Quebec AI Institute 3Element AI / Service Now, 4Independent Robotics, 5Algolux, 6Mc Gill University, 7École de technologie supérieure, 8Canada CIFAR AI Chair, 9Princeton University.
Pseudocode No The paper does not contain any block explicitly labeled 'Pseudocode' or 'Algorithm'.
Open Source Code No The paper mentions 'We implemented our model using the Pytorch open-source framework.' but this refers to a third-party framework, not the authors' own code for the described methodology. While a URL to a project page is provided, it does not explicitly state that the source code for the paper's methodology is available there.
Open Datasets Yes We evaluate Auto Bot on nu Scenes (Caesar et al., 2020) and Argoverse (Chang et al., 2019), two autonomous driving trajectory prediction benchmarks, on the synthetic partition of the Traj Net++ (Kothari et al., 2021) dataset, a pedestrian trajectory forecasting benchmark, and Omniglot (Lake et al., 2015), a dataset of hand-drawn characters that we use for predicting strokes.
Dataset Splits Yes There are a total of 54, 513 unique scenes in the dataset, which we split into 49, 062 training scenes and 5, 451 validation scenes.
Hardware Specification Yes A distinguishing feature of Auto Bots is that all models are trainable on a single desktop GPU (1080 Ti) in under 48h. ... Auto Bot-Ego was trained on a single Nvidia Geforce GTX 1080Ti GPU, using approximately 2 GB of VRAM.
Software Dependencies No The paper states 'We implemented our model using the Pytorch open-source framework,' but does not specify version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes Table 4 shows the values of parameters used in our experiments across the different datasets. ... For the Nuscenes and Traj Net++ datasets, we anneal the learning rate every 10 epochs for the first 20 epochs by a factor of 2, followed by annealing it by a factor of 1.33 every 10 epochs for the next 30 epochs. ... we found it helpful to clip the gradients to a maximum magnitude of 5.0.