Neurosim

Blazing fast multirotor simulator with event camera support. Pythonic and real-time.

Richeek Das, Pratik Chaudhari

GRASP Laboratory, University of Pennsylvania

[📜 Paper] • [📖 BibTeX]

Quick Start: If you only need a fast CUDA event simulator, we've made it standalone. Learn how to use our optimized event simulator at grasp-lyrl/neurosim_cu_esim.

📰 News

• Hardware-in-the-loop (HIL) support for real robot testing
• 📚 Detailed documentation and tutorials
• 06/21/26: Added online dataloader with shared mem for Multi GPU simulation and learning.
• 04/16/26: Added dynamic obstacles thrown at the agent. See GIF for fun.
• 03/30/26: Reinforcement Learning (RL) support added.
• 02/16/26: Neurosim paper released on arXiv (2602.15018).
• 02/15/26: neurosim_cu_esim now standalone for faster event simulation.
• 02/06/26: Initial release of Neurosim!

📋 Table of Contents

• Applications
• Installation
  • Option 1: Docker (Recommended)
  • Option 2: Conda
• Basic Usage
• Citation
• Issues

Applications

Neurosim enables real-time closed-loop control (src/neurosim/sims/asynchronous_simulator, src/neurosim/cortex/), online training of multi-modal perception models (applications/f3_training/), and event-based reinforcement learning with NeurosimRL (applications/rl/). Detailed docs to run each of them are coming soon!

Rollout with a trained policy to stabilize a "thrown" quad (without crashing into things)

Dynamic obstacles thrown at the quad

RL Quick Start (reproducing the cool stuff above)

Download the pretrained checkpoint folder from here and place them inside applications/rl/example_ckpt/hover_sb3_combined_experiment

Then (once you have Neurosim installed) you can rollout the trained policy with:

python applications/rl/run_policy.py \
  --checkpoint applications/rl/example_ckpt/hover_sb3_combined_experiment/best_model.zip \
  --rollout-config applications/rl/configs/hover_sb3_combined_rollout.yaml \
  --episodes 10 \
  --visualize

Installation

Option 1: Docker (Recommended)

Docker provides a consistent environment with all dependencies pre-configured, including CUDA and ZMQ libraries. Two image variants are available:

• ros (default): includes ROS2 Humble and ROS development tooling
• noros: excludes ROS2 for a lighter image when ROS integration is not needed

Prerequisites

• Docker with NVIDIA GPU support (nvidia-docker)
• NVIDIA drivers installed on host (nvcc 12.9+)

Pull Prebuilt Images (Recommended)

# Clone the repository
git clone https://github.com/grasp-lyrl/neurosim.git
cd neurosim

# Pull ROS-enabled image
docker pull richeek01/neurosim:ros

# OR Pull image without ROS2
docker pull richeek01/neurosim:noros

# Tag pulled images so existing run script works unchanged
docker tag richeek01/neurosim:ros neurosim:ros
docker tag richeek01/neurosim:noros neurosim:noros
docker tag richeek01/neurosim:latest neurosim:latest

Build Locally (Optional)

# Build ROS-enabled image (default, takes ~15-20 minutes)
bash docker/build.sh ros

# Build image without ROS2
bash docker/build.sh noros

Run the Container

# Launch ROS-enabled container (default)
bash docker/run.sh ros

# Launch container without ROS2
bash docker/run.sh noros

This will:

• Mount the current directory to /home/${USER}/neurosim inside the container
• Enable GPU access with all CUDA capabilities
• Forward X11 display for GUI applications
• Set up shared networking and IPC

Inside the Container

# Navigate to the workspace
cd neurosim

# neurosim environment should be activated by default. If not, run:
conda activate neurosim

# Download example scenes
python -m habitat_sim.utils.datasets_download --uids habitat_test_scenes --data-path data/

# Run the simulator
python test_sim.py --settings configs/skokloster-castle-settings.yaml --display --world_rate 750

---

Option 2: Conda

System Requirements

• OS: Ubuntu 22.04/24.04 (tested)
• Compiler: GCC 11.4.0+
• CMake: 3.14.0+
• CUDA: 12.2 / 12.4 / 12.6 / 12.9 / 13.0 (tested)
• Python: 3.10

Install System Dependencies

sudo apt-get update
sudo apt-get install -y --no-install-recommends \
    libjpeg-dev libglm-dev libgl1-mesa-glx libegl1-mesa-dev \
    mesa-utils xorg-dev freeglut3-dev

Install Neurosim

# Create conda environment
conda create -n neurosim python=3.10 cmake=3.14.0 pip==25.1.1 -y
conda activate neurosim

# Install neurosim in editable mode
pip install -e . -v

---

Basic Usage

Download Example Data

python -m habitat_sim.utils.datasets_download --uids habitat_test_scenes --data-path data/

Run the Simulator

python test_sim.py --settings configs/skokloster-castle-settings.yaml --display

Citation

If you use this code in your research, please cite:

@misc{das2026neurosim,
      title={Neurosim: A Fast Simulator for Neuromorphic Robot Perception}, 
      author={Richeek Das and Pratik Chaudhari},
      year={2026},
      eprint={2602.15018},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2602.15018}, 
}

Issues

Please report any bugs or feature requests on GitHub issues. Pull requests are very welcome!

Compilation Issues

• If compilation crashes due to high memory usage or CPU load, manually set self.parallel=4 inside setup.py of habitat-sim to limit parallel jobs.

• pip==25.3 breaks installation due to changes in the build isolation process. Use pip==25.1.1 as specified.

License

Apache 2.0 License - see the LICENSE file for details.

Acknowledgements

This project is enabled by the following amazing open-source projects: Habitat-Sim, ZeroMQ, RotorPy

Banner credits: @ongdexter, Gemini 3