Conceptual visualization of AEGIS AI-driven thrust vector guidance
01Mission Statement
Traditional rocket guidance systems use PID controllers — proven, reliable, and deeply boring. They follow pre-programmed trajectories and react to disturbances with fixed mathematical relationships that were tuned on the ground. AEGIS asks: what if the rocket could learn to fly itself?
Using reinforcement learning trained in high-fidelity simulation, AEGIS develops control policies that can handle wind gusts, thrust asymmetries, mass distribution changes, and off-nominal engine performance — all in real-time, all without pre-programmed response tables. The result is a guidance system that doesn't just follow a trajectory; it discovers the optimal one.
02Core Technology
Reinforcement Learning Agent
Our RL agent uses Proximal Policy Optimization (PPO) trained in a Simulink-based rocket dynamics simulator. The agent observes IMU data (accelerometer, gyroscope), GPS position (when available), barometric altitude, and thrust telemetry — and outputs gimbal commands for the thrust vector control actuators at 100Hz.
6-DOF Simulation Environment
Our training environment models full 6-degree-of-freedom rocket dynamics including aerodynamic forces, gravity gradient, atmospheric density profiles, wind models (Dryden turbulence), and thrust curve variations. Domain randomization ensures the agent generalizes across a wide range of conditions.
Thrust Vector Control Hardware
AEGIS targets a gimbaled nozzle mechanism with sub-degree pointing accuracy and <100ms response time. The mechanical design uses dual-axis servo-driven gimbal rings rated for the thermal environment downstream of the IGNIS motor. Currently in CAD design phase with FEA structural analysis.
ROS2 Integration Layer
The entire AEGIS stack runs on ROS2, enabling modular node architecture. Sensor drivers, state estimation (EKF), the RL policy inference, and actuator commands all communicate via low-latency DDS transport. This allows hardware-in-the-loop testing by swapping simulated nodes for real hardware incrementally.
03Training Results
04Convergence with IGNIS
"The best guidance system is one that learns from every flight — including the ones that went wrong. Especially the ones that went wrong."
AEGIS is designed to eventually integrate with the IGNIS motor. The gimbal mechanism mounts directly to the IGNIS nozzle assembly, and the control system receives thrust telemetry from IGNIS's DAQ system. Our roadmap has hardware-in-the-loop testing with the IGNIS motor by Q3 2026, and a first guided flight test on a sounding rocket by Q1 2027.
The combination of AEGIS + IGNIS represents JPL's path to a fully autonomous, AI-guided launch vehicle — one that can adapt its flight plan in real-time based on actual conditions, not pre-programmed assumptions.
05Control Validation
Classical Baseline
AEGIS policies are compared against tuned PID and model-predictive control baselines. Reinforcement learning only earns its place when it improves recovery behavior without sacrificing predictable stability.
Safety Envelope
The learned policy is wrapped in limit checks for gimbal angle, actuator rate, attitude error, dynamic pressure, and sensor validity. Guidance authority can be reduced or rejected if the vehicle leaves the approved envelope.
Hardware-in-the-Loop
Flight software, IMU streams, actuator commands, and simulated vehicle dynamics are tested together before any motor integration. This catches timing, filtering, and command saturation issues early.
Abort Logic
AEGIS is designed with conservative abort states. If estimation confidence collapses or actuator response diverges, the system favors a known safe behavior over continued optimization.
06Development Roadmap
The immediate path is simulation maturity: better aerodynamic models, actuator lag, noisy sensors, and randomized thrust curves. The next step is hardware-in-the-loop validation with the same embedded stack intended for flight tests.
AEGIS becomes meaningful when it closes the loop with real propulsion data. IGNIS provides measured thrust behavior, NOVA can inform perception and state awareness, and AEGIS turns those signals into controlled flight decisions.
NEXTExplore Next
Interested in AEGIS?
We're looking for control systems engineers, RL researchers, and simulation specialists.