<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> This project addresses the challenge of safe and intention-aware robot navigation in dense, interactive crowds. It targets researchers and engineers by providing a novel recurrent graph neural network with attention mechanisms that models heterogeneous agent interactions and predicts human intentions. This approach enables robots to navigate effectively and non-invasively, preventing conflicts by considering others' future paths, and has been demonstrated in simulation and real-world robot deployment.

How It Works

The core of CrowdNav++ is a recurrent graph neural network augmented with attention mechanisms. This architecture captures complex, heterogeneous interactions among agents across space and time. A key innovation is the prediction of dynamic agents' future trajectories to infer their intentions. These predictions are then integrated into a model-free reinforcement learning framework, guiding the robot to avoid intruding into the intended paths of other agents, thereby enhancing navigation safety and efficiency.

Quick Start & Requirements

Setup requires a Python 3.x environment. Install dependencies via pip install -r requirements.txt, followed by PyTorch 1.12.1, OpenAI Baselines, and the Python-RVO2 library. Training involves modifying configuration files (crowd_nav/configs/config.py, arguments.py) and running python train.py. Testing is initiated by adjusting test.py parameters and executing python test.py. Links to project website, arXiv, and YouTube demonstrations are provided.

Highlighted Details

• Successfully demonstrated sim-to-real transfer of learned policies to a TurtleBot 2i.
• Integrates a Gumbel Social Transformer (GST) for human trajectory prediction.
• Offers pre-trained models for various scenarios.
• Supports visualization of predicted trajectories and training curves.

Maintenance & Community

The project is associated with ICRA 2023 and lists several key contributors. Follow-up work, including the HEIGHT model and a sim2real tutorial, is available. For questions or bug reports, users are encouraged to open issues or pull requests.

Licensing & Compatibility

The repository's README does not specify a software license. Compatibility is confirmed for Ubuntu with Python 3.6 and 3.8, with no guarantees for other operating systems or Python versions.

Limitations & Caveats

The codebase does not include training scripts for the trajectory prediction network, referencing an external repository. Performance may vary due to hyperparameters and random seeds, necessitating manual tuning. Sim-to-real transfer reproducibility is not guaranteed due to real-world uncertainties. Explicit license information is absent.