[Paper Review] Flightmare: A Flexible Quadrotor Simulator

Yunlong Song, Selim Naji+3 more|arXiv (Cornell University)|Sep 1, 2020

Robotic Path Planning Algorithms23 references58 citations

TL;DR

Flightmare decouples rendering and physics to provide a fast, configurable quadrotor simulator with rich sensors, multi-agent RL support, and VR integration. It achieves up to 230 Hz rendering and 200,000 Hz dynamics on a laptop and supports hundreds of drones in parallel.

ABSTRACT

State-of-the-art quadrotor simulators have a rigid and highly-specialized structure: either are they really fast, physically accurate, or photo-realistic. In this work, we propose a novel quadrotor simulator: Flightmare. Flightmare is composed of two main components: a configurable rendering engine built on Unity and a flexible physics engine for dynamics simulation. Those two components are totally decoupled and can run independently of each other. This makes our simulator extremely fast: rendering achieves speeds of up to 230 Hz, while physics simulation of up to 200,000 Hz on a laptop. In addition, Flightmare comes with several desirable features: (i) a large multi-modal sensor suite, including an interface to extract the 3D point-cloud of the scene; (ii) an API for reinforcement learning which can simulate hundreds of quadrotors in parallel; and (iii) integration with a virtual-reality headset for interaction with the simulated environment. We demonstrate the flexibility of Flightmare by using it for two different robotic tasks: quadrotor control using deep reinforcement learning and collision-free path planning in a complex 3D environment.

Motivation & Objective

• Enable fast and physically flexible quadrotor simulation by decoupling rendering from dynamics.
• Provide a rich sensor suite and environment options for robotics and RL research.
• Support large-scale parallel data collection and training for learning-based control.
• Bridge the gap between simulation and real-world robotics with VR and real-time interfaces.

Proposed method

• Two decoupled blocks: a Unity-based rendering engine and a flexible quadrotor dynamics model.
• Asynchronous ZeroMQ-based interface enables fast communication between rendering and dynamics blocks.
• Three dynamics options: Gazebo-based, real-world, and parallelized classical quadrotor dynamics.
• Rendering supports configurable sensors (RGB, depth, segmentation) and camera effects; outputs ground-truth point clouds in PLY via Open3D.
• OpenAI Gym-style Python API enables RL training; tasks use PPO with parallel sampling (up to ~2M samples per minute at scale).

Experimental results

Research questions

• RQ1 How does decoupling rendering and physics affect simulation speed and usability for learning-based control?
• RQ2 Can Flightmare accelerate data collection for RL by simulating hundreds of quadrotors in parallel?
• RQ3 What sensor modalities and environment features are most helpful for learning and perception research?
• RQ4 How effectively can the simulator support path planning, RL tasks, and VR-based human-robot interaction?

Key findings

• Rendering can exceed real-time performance, up to 230 Hz in Garage environments with multiple cameras.
• Physics dynamics can reach up to 200,000 steps per second on a multi-core CPU.
• 150 quadrotors in parallel achieve over 200,000 steps/s; ~2 million samples per minute with an MLP policy and PPO.
• RL tasks include quadrotor stabilization, motor-failure recovery, and fast gate traversal, trained with PPO over 25M timesteps per task.
• Flightmare exports environment point clouds (PLY) for planning with OMPL, enabling 1-second shortest-path computation in forest environments.
• The VR integration enables safe, collocated human-robot interaction experiments.