yyf17 / social-nce-crowdnav

Official implementation of the "Social NCE: Contrastive Learning of Socially-aware Motion Representations" in PyTorch.

CrowdNav with Social-NCE

This is an official implementation for the paper

Social NCE: Contrastive Learning of Socially-aware Motion Representations
by Yuejiang Liu, Qi Yan, Alexandre Alahi at EPFL

TL;DR: Contrastive Representation Learning + Negative Data Augmentations 🡲 Robust Neural Motion Models

• Rank in 1st place on the Trajnet++ challenge since November 2020 to present
• Significantly reduce the collision rate of SOTA human trajectroy forecasting models
• SOTA on imitation / reinforcement learning for autonomous navigation in crowds

Please check out our code for experiments on different models as follows:
Social NCE + CrowdNav | Social NCE + Trajectron | Social NCE + STGCNN

Preparation

Setup environments follwoing the SETUP.md

Training & Evaluation

• Behavioral Cloning (Vanilla)

  python imitate.py --contrast_weight=0.0 --gpu
  python test.py --policy='sail' --circle --model_file=data/output/imitate-baseline-data-0.50/policy_net.pth
  

• Social-NCE + Conventional Negative Sampling (Local)

  python imitate.py --contrast_weight=2.0 --contrast_sampling='local' --gpu
  python test.py --policy='sail' --circle --model_file=data/output/imitate-local-data-0.50-weight-2.0-horizon-4-temperature-0.20-nboundary-0-range-2.00/policy_net.pth
  

• Social-NCE + Safety-driven Negative Sampling (Ours)

  python imitate.py --contrast_weight=2.0 --contrast_sampling='event' --gpu
  python test.py --policy='sail' --circle --model_file=data/output/imitate-event-data-0.50-weight-2.0-horizon-4-temperature-0.20-nboundary-0/policy_net.pth
  

• Method Comparison

  bash script/run_vanilla.sh && bash script/run_local.sh && bash script/run_snce.sh
  python utils/compare.py
  

Basic Results

Results of behavioral cloning with different methods.

Averaged results from the 150th to 200th epochs.

	collision	reward
Vanilla	12.7% ± 3.8%	0.274 ± 0.019
Local	19.3% ± 4.2%	0.240 ± 0.021
Ours	2.0% ± 0.6%	0.331 ± 0.003

Citation

If you find this code useful for your research, please cite our papers:

@article{liu2020snce,
  title   = {Social NCE: Contrastive Learning of Socially-aware Motion Representations},
  author  = {Yuejiang Liu and Qi Yan and Alexandre Alahi},
  journal = {arXiv preprint arXiv:2012.11717},
  year    = {2020}
}
@inproceedings{chen2019crowdnav,
    title={Crowd-Robot Interaction: Crowd-aware Robot Navigation with Attention-based Deep Reinforcement Learning},
    author={Changan Chen and Yuejiang Liu and Sven Kreiss and Alexandre Alahi},
    year={2019},
    booktitle={ICRA}
}