TL;DR: We propose a framework for obtaining multimodal distributions over the future evolution of elements in sets based on latent variables and Transformer-based modelling of sequences. Based on the framework, we introduce a class of model architectures called AutoBots that use multi-head self-attention blocks for social attention between elements and for modelling the time-evolution of individual elements in the set. With the use of learnable seed parameters in the decoder, AutoBots are computationally efficient and can be used for real-time inference. We validate AutoBots in the context of motion prediction for autonomous driving where we achieve strong results on the Nuscenes dataset. We further demonstrate its multi-agent forecasting performance on a synthetic pedestrian dataset.
Abstract
Robust multi-agent trajectory prediction is essential for the safe control of robotic systems. A major challenge is to efficiently learn a representation that approximates the true joint distribution of contextual, social, and temporal information to enable planning. We propose Latent Variable Sequential Set Transformers which are encoder-decoder architectures that generate scene-consistent multi-agent trajectories. We refer to these architectures as “AutoBots”. The encoder is a stack of interleaved temporal and social multi-head self-attention (MHSA) modules which alternately perform equivariant processing across the temporal and social dimensions. The decoder employs learnable seed parameters in combination with temporal and social MHSA modules allowing it to perform inference over the entire future scene in a single forward pass efficiently. AutoBots can produce either the trajectory of one ego-agent or a distribution over the future trajectories for all agents in the scene. For the single-agent prediction case, our model achieves top results on the global nuScenes vehicle motion prediction leaderboard, and produces strong results on the Argoverse vehicle prediction challenge. In the multi-agent setting, we evaluate on the synthetic partition of TrajNet++ dataset to showcase the model’s socially-consistent predictions. We also demonstrate our model on general sequences of sets and provide illustrative experiments modelling the sequential structure of the multiple strokes that make up symbols in the Omniglot data. A distinguishing feature of AutoBots is that all models are trainable on a single desktop GPU (1080 Ti) in under 48h.
Paper
The paper can be found on arXiv.org: https://arxiv.org/abs/2104.00563
The work was originally published on 19th Feb 2021 and got accepted at ICLR 2022 as spotlight
Video/poster will follow soon.
Code
The full repo is available in Pytorch here.
Examples
Nuscenes
On the challenging NuScenes dataset, our model learns plausible paths through intersections. Top row: intersection birdseye view with ground truth input (cyan) and output (magenta) trajectory. Bottom row: learned plausible paths for the agent.
TrajNet++ Synthetic Data
The synthetic partition of the TrajNet++ dataset is designed to include a high-level of interaction between agents. The top-row shows the performance of our full model (labelled “AutoBot”) which predicts scene-consistent and plausible future scenes across its 3 different modes. We contrast this to two other AutoBot variants which do not employ social attention in the decoder (middle-row, AutoBot-Ego), or in both the encoder and decoder (bottom-row, AutoBot-Antisocial).
Computational Efficiency
Conventional methods employ autoregressive decoding to generate the future scene which can make such methods slow at inference time. One of the main contributions of AutoBots is the use of learnable seed parameters in the decoder which allows it to predict the entire future scene with a single forward pass. Here we show how increasing the number of agents, the number of input timesteps, the prediction horizon and the number of discrete modes affect the inference speed (in FPS) of AutoBot-Ego (top-row) and AutoBot (bottom-row). We can see that models with the decoder seed parameters can perform real-time inference (>30FPS) for almost all configurations, while autoregressive counterparts struggle depending on the configuration. All runs were performed on a single desktop GPU (1080 Ti).
Bibtex
@inproceedings{
girgis2022latent,
title={Latent Variable Sequential Set Transformers for Joint Multi-Agent Motion Prediction},
author={Roger Girgis and Florian Golemo and Felipe Codevilla and Martin Weiss and Jim Aldon D'Souza and Samira Ebrahimi Kahou and Felix Heide and Christopher Pal},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=Dup_dDqkZC5}
}