Aerial Gym Simulator — Isaac Lab Edition

Welcome to the documentation of the Aerial Gym Simulator (Isaac Lab Edition)   

The Aerial Gym Simulator is a high-fidelity, GPU-accelerated simulator for training Micro Aerial Vehicle (MAV) platforms to learn to fly using reinforcement learning. This fork has been fully ported to NVIDIA Isaac Lab (Isaac Sim 4.5.0), replacing the deprecated Isaac Gym backend. We offer aerial robot models for standard quadrotor platforms (base quad, x500, lmf2), each with a dedicated reinforcement learning environment and reward function. Policies train in minutes on a single GPU and can be visualized live in the Isaac Sim GUI.

This edition is open-source and released under the BSD-3-Clause License.

Train a hovering policy in minutes and watch it converge in the Isaac Sim GUI:

GPU-parallelized training across thousands of environments simultaneously:

Features

• Isaac Lab Native

  Fully ported to Isaac Lab 2.1.0 — no Isaac Gym required. Uses DirectRLEnv for clean, minimal environment definitions with full Isaac Sim 4.5.0 compatibility.
• 5 Ready-to-Train Environments

  PositionSetpointEnv, PositionSetpointSim2RealEnv, PositionSetpointEndToEndEnv, PositionSetpointX500Env, PositionSetpointLMF2Env — each with a dedicated PPO config and play script.
• skrl PPO Training

  Uses skrl 1.4.3 with RunningStandardScaler observation normalization, fixed learning rate, and deterministic eval mode. Train headless with --num_envs 4096 for fastest convergence.
• Testable Reward Functions

  Reward functions are extracted into standalone @torch.jit.script modules (sim2real_reward.py, end_to_end_reward.py, lmf2_reward.py) that can be imported and unit-tested without launching Isaac Sim.
• Rotation Utilities

  rotation_utils.py provides quaternion ↔ rotation matrix ↔ ZYX Euler conversions, 6D continuous rotation representation for end-to-end policies, and small-angle wrapping — all JIT-compiled.
• Sim2Real Ready

  Observation noise (position ±0.03 m, orientation ±0.02 rad, velocity ±0.02 m/s) baked into PositionSetpointSim2RealEnv and lmf2. Asymmetric closer_reward in lmf2 penalizes retreating 3× harder than approaching.
• Unit Tests — No SimApp Needed

  14 tests covering all reward functions and rotation utilities. Run in ~1.5 s on CPU with no GPU and no running simulator, using a custom import hook that stubs omni.* and isaaclab.sim automatically.

Why Aerial Gym Simulator — Isaac Lab Edition?

Isaac Gym was deprecated by NVIDIA. This edition migrates the full simulator to Isaac Lab, the supported successor, while keeping the same simple training interface. You get:

• One command to train — train_skrl_position_setpoint.py --num_envs 4096 --headless
• One command to visualize — play_skrl_position_setpoint.py --checkpoint best_agent.pt --num_envs 16
• Fast iteration — reward functions test in 1.5 s without launching the simulator
• Three robot models — base quad, x500 (1.656 kg), lmf2 (1.24 kg, base-link wrench control)
• Checkpoint resume — training interrupted? Resume from any saved checkpoint

Policies converge in 5M environment steps (~2–4 h on a modern 24 GB GPU with 4096 envs).

Environments

Environment	Obs	Robot	Notes
PositionSetpointEnv	13	base quad	simplest hovering task
PositionSetpointSim2RealEnv	17	base quad	sensor noise, sim2real reward
PositionSetpointEndToEndEnv	15	base quad	rotation 6D obs
PositionSetpointX500Env	15	x500 (1.656 kg)	per-motor force, 6D rotation
PositionSetpointLMF2Env	17	lmf2 (1.24 kg)	base-link wrench, asymmetric reward

Quick Links

• Installation
• Train a hovering policy
• All environments and training scripts
• Reward function design
• Unit tests
• FAQ & Troubleshooting

Citing

When referencing the Aerial Gym Simulator in your research, please cite the original paper:

@ARTICLE{kulkarni2025aerial,
  author={Kulkarni, Mihir and Rehberg, Welf and Alexis, Kostas},
  journal={IEEE Robotics and Automation Letters},
  title={Aerial Gym Simulator: A Framework for Highly Parallelized Simulation of Aerial Robots},
  year={2025},
  volume={10},
  number={4},
  pages={4093-4100},
  doi={10.1109/LRA.2025.3548507}
}

Contact

Ted Lin      GitHub

For bugs and feature requests, please open an Issue on GitHub.

The original Aerial Gym Simulator was developed at the Autonomous Robots Lab, Norwegian University of Science and Technology (NTNU).

Acknowledgements

This Isaac Lab edition builds on the original Aerial Gym Simulator by Mihir Kulkarni, Welf Rehberg, and Kostas Alexis (NTNU ARL). Original work supported by RESNAV (AFOSR Award No. FA8655-21-1-7033) and SPEAR (Horizon Europe Grant Agreement No. 101119774).

FAQs and Troubleshooting

See the FAQ page or open an Issue on GitHub.