icarus
Autonomous Rocket Landing. RL-trained 6-DOF rocket simulation with curriculum learning and real-time telemetry.
ICARUS trains reinforcement learning agents to master rocket propulsive landing through curriculum learning and domain randomization.
Initializing simulation...
Mission Gallery
From hover to full suborbital hop — every stage visualized.
Precision Landing
Watch AI agents nail propulsive landings from any entry angle, altitude, or velocity.
Curriculum Mastery
5-stage training from hover to full suborbital hop with progressive difficulty.
Live Telemetry
Real-time altitude, velocity, throttle, attitude, and neural activity visualization.
Multi-Agent GNC
Four specialized agents coordinate guidance, navigation, control, and safety.
ICARUS (Intelligent Control for Autonomous Rocket Upright Stabilization) is a full-fidelity rocket flight simulation with AI-powered guidance, navigation, and control. The system uses PPO/SAC reinforcement learning with a 5-stage curriculum to train agents that achieve 95%+ landing success rates.
The physics engine runs at 1kHz with RK4 numerical integration, modeling US Standard Atmosphere 1976, Dryden wind turbulence, J2 gravity perturbations, and full 6-DOF rigid body dynamics. Four specialized agents handle GNC, attitude control, throttle management, and safety monitoring.
"The future of autonomous flight is training in simulation and transferring to reality. ICARUS proves that reinforcement learning can solve the hardest control problems in aerospace when given the right curriculum and physics fidelity."
The Philosophy
ICARUS isn't just a simulation — it's a proving ground for the thesis that RL can solve safety-critical control problems when paired with the right curriculum.
Sim-to-Real Transfer
The gap between simulation and reality is the central challenge of robotics RL. ICARUS addresses this with domain randomization — varying wind, gravity, mass properties, and sensor noise across training episodes so the agent learns policies robust to the real-world variations it will encounter. The physics engine's fidelity (US Std Atmosphere 1976, Dryden turbulence, J2 gravity) ensures the simulation is close enough to reality that transfer is possible.
Why Curriculum Learning Works
You wouldn't teach a student calculus before algebra. Similarly, ICARUS uses a 5-stage curriculum that builds mastery incrementally: Stage 1 (Hover) teaches basic thrust control. Stage 2 (Vertical Drop) adds altitude management. Stage 3 (Offset Landing) introduces lateral correction. Stage 4 (High Entry) adds full trajectory planning. Stage 5 (Full Mission) combines everything into a complete suborbital hop. Each stage's reward function shapes behavior toward the next stage's starting conditions.
Multi-Agent GNC Architecture
Rather than training a single monolithic policy, ICARUS decomposes the control problem into four specialized agents: GNC (trajectory planning and guidance), ATT (attitude control and stabilization), THR (throttle management and fuel optimization), and SAF (safety monitoring and abort decisions). Each agent trains on its specific subtask while communicating through a shared observation space. This decomposition mirrors real aerospace GNC architectures and produces more interpretable, more robust policies.
Kingly's Approach
We built ICARUS as a full-stack RL platform combining aerospace-grade physics with modern deep RL training infrastructure.
1kHz Physics Engine
The core simulation runs at 1000Hz with RK4 numerical integration for stability. We model drag, gravity (with J2 perturbation), thrust vectoring, mass flow dynamics, atmospheric density variation, and wind turbulence. This fidelity is essential — agents trained on simplified physics fail when exposed to real-world conditions.
Training Infrastructure
ICARUS uses Stable-Baselines3 for PPO and SAC training with custom gymnasium environments. Training runs are managed via Weights & Biases for experiment tracking, with automatic curriculum advancement based on rolling success rate thresholds. A full training run from Stage 1 to Stage 5 takes approximately 48 hours on a single A100 GPU.
Real-Time 3D Dashboard
The visualization layer uses Three.js for 3D rocket rendering with real-time telemetry overlays. The dashboard shows altitude, velocity, throttle, attitude, neural activity patterns, mission phase progress, and agent status — all updated at 60fps from the simulation backend via WebSocket.
The Future
Autonomous landing is just the beginning. The simulation platform extends to multiple aerospace control problems.
Atmospheric Re-entry
Extending the simulation to model full orbital re-entry with ablative heat shielding, plasma effects, and skip trajectories.
Multi-Vehicle Coordination
Training swarms of rockets to perform coordinated landings, formation flying, and cooperative docking maneuvers.
Hardware-in-the-Loop
Connecting the simulation to actual flight computers and actuators for pre-flight validation of trained policies.
Lunar/Mars Landing
Adapting the physics and training curriculum for reduced-gravity environments with different atmospheric conditions.
Tech Stack
What This Is Used For
Propulsive landing R&D for reusable launch vehicles
RL curriculum design for safety-critical control domains
High-fidelity aerospace simulation and visualization
Multi-agent GNC system prototyping
Install
NPM Package Coming Soon
Sign up for the newsletter to get notified when it's released.
Also from Labs
Stay Updated
Get notified of new labs, experiments, and updates.