OmniDrones Documentation

OmniDrones is an open-source platform for reinforcement learning (RL) research on multi-rotor drone systems. It is built on top of Nvidia Isaac Sim and provides a highly efficient and flexible simulation environment. The platform includes a suite of benchmark tasks and algorithm baselines to assist in drone control research.

Key Requirements & Compatibility:

• Simulator: Built on Nvidia Isaac Sim (current version uses 4.1.0).
• Python Version: Requires Python 3.10.
• OS Support: Does not support Windows (use Linux/Ubuntu).
• Documentation: Detailed usage instructions are available at omnidrones.readthedocs.io.

OmniDrones provides several multi-agent task environments. These tasks are categorized by the type of agent and the objective:

Platform-based Tasks:

• PlatformHover: Hovering over a platform.
• PlatformTrack: Tracking a platform.
• PlatformFlyThrough: Flying through a platform-based environment.

Transport-based Tasks:

• TransportHover: Hovering near/over transport objects.
• TransportTrack: Tracking transport objects.
• TransportFlyThrough: Flying through transport-based environments.

Formation Tasks:

• Formation: Coordinating multiple agents to maintain a specific formation.

OmniDrones provides several single-agent reinforcement learning tasks. These tasks are categorized by their objective and environment complexity:

Basic Navigation Tasks

• Hover: Maintaining a stable position.
• Track: Following a specific trajectory or target.
• FlyThrough: Navigating through a series of waypoints or obstacles.

Payload-Related Tasks

These tasks involve interacting with or carrying an external payload:

• PayloadHover
• PayloadTrack
• PayloadFlyThrough

Inverted Pendulum Tasks

Tasks focused on stabilization and control dynamics:

• InvPendulumHover
• InvPendulumTrack
• InvPendulumFlyThrough

Complex Environments

• Forest: Navigation through obstacle-dense environments (often requires LiDAR).
• Pinball: Dynamic obstacle avoidance or target interaction.

The reward is a composite of several factors designed to encourage trajectory tracking, safety, and stability. The total reward $r$ is calculated as:

$r = r_{\text{separation}} \times (r_{\text{pos}} + r_{\text{pos}} \times (r_{\text{up}} + r_{\text{swing}}) + r_{\text{joint_limit}})$

Individual reward components include:

• pos: Trajectory tracking reward, computed as $\exp(-a \times \text{pos_error})$.
• separation: A penalty applied to all agents when the minimum separation between them is too small.
• up: Reward for maintaining the payload in an upright position.
• swing: Reward for minimizing large swinging motions of the payload.
• joint_limit: Reward for keeping joint states within a range that avoids glitchy behaviors.
• effort: Reward based on drone effort to optimize energy consumption.

An episode in PayloadTrack ends when any of the following occur:

1. The drone gets too close to the ground.
2. The distance between the payload and the target exceeds the reset_thres threshold.
3. The maximum episode length is reached.