iRonCub-Mk3: Jet-Powered Aerial Humanoid

Updated 3 July 2026

iRonCub-Mk3 is a jet-powered floating-base humanoid robot that integrates thrust vectoring and agile whole-body dynamics for controlled aerial mobility.
It employs advanced aerodynamic modeling, model predictive control, and sensor fusion to achieve robust vertical takeoff and accurate trajectory tracking.
A CAD-driven co-design approach optimizes its structural integrity and mechanical performance, ensuring safe operations with high thrust-to-weight ratios.

iRonCub-Mk3 is a jet-powered floating-base humanoid robot built on the iCub3 platform and engineered for controlled aerial mobility using four custom-mounted turbojet engines. It constitutes one of the central research platforms for the study of humanoid flight, enabling an unprecedented integration of whole-body robot dynamics, aerodynamic modeling, advanced Model Predictive Control (MPC), and CAD-driven mechanical co-design. iRonCub-Mk3’s development and its first sustained vertical takeoff represent a salient step toward flight-ready aerial humanoids (Gorbani et al., 1 Jun 2025).

1. Physical Architecture and Propulsion System

iRonCub-Mk3 retains the kinematic structure of iCub3, comprising a 6-DoF free-floating base and 45 actuated joints spanning the head, torso, arms, and legs— $q \in \mathbb{R}^{6+45}$ . The propulsion system consists of four JetCat P250 Pro small-scale turbojets: two are integrated into a detachable jetpack affixed to the torso and two are mounted on forearm brackets below the elbows (Gorbani et al., 1 Jun 2025). Each engine is capable of peak static thrust ≈250 N, for a system-level thrust-to-weight ratio $T/W\approx3.4$ (robot mass $m\approx30$ kg).

Engine mounting geometry is selected to balance thrust efficacy, control authority, and thermal management; thrust axes are angled slightly (≈5–10° tilt) to optimize both attitude control and minimize structural exposure to ≈600 °C exhaust (Vanteddu et al., 2024). Jetpack and forearm brackets employ high-strength steel or titanium; the robot core remains an aluminum alloy. Finite Element Method (FEM) analyses with 750 N axial loads (3×max thrust) ensure minimum structural safety factors above 2.5 for primary jet interfaces. Heat shielding with aerogel is employed at critical locations (Gorbani et al., 1 Jun 2025).

A tabulation of key mechanical subsystems is provided:

Subsystem	Material(s)	Safety Factor (FEM)	Unique Features
Core structure	Aluminum alloy	—	iCub3 heritage
Jetpack bracket	Steel/Titanium	>2.5	Custom orientation; FEM stress benchmarked
Forearm support	Steel/Titanium	>2.5	Adjustable geometry for thrust vector tuning
Aerogel shielding	Silica composite	—	Thermal insulation ( $\approx$ 600 °C)

2. Aerodynamic and Multibody Modeling

The dynamics are modeled using full Euler–Lagrange joint-space equations,

$H(q)\,\ddot{q} + C(q,\dot{q})\,\dot{q} + g(q) = \tau + J^\top F_{\rm ext},$

combined with centroidal momentum dynamics,

$\dot h = A(q)T + mge_3,$

where $h \in \mathbb{R}^6$ are the total linear/angular momenta, $A(q)$ is the actuator-to-centroidal wrench map, $T \in \mathbb{R}^4$ is the vector of jet thrusts, and $e_3$ is the gravity axis.

Aerodynamic modeling incorporates Computational Fluid Dynamics (CFD)-based drag and lift maps parameterized over angle of attack ( $T/W\approx3.4$ 0) and sideslip ( $T/W\approx3.4$ 1), with force components $T/W\approx3.4$ 2, $T/W\approx3.4$ 3 analytically fit to the results of 45 steady RANS scenarios. The drag and normal force coefficients are mapped as

$T/W\approx3.4$ 4

$T/W\approx3.4$ 5

with coefficients calibrated on CFD data for flight envelope control (Hui et al., 2022).

These aerodynamic components enter both the floating-base and centroidal equations, allowing for simulation and control of wind- and attitude-dependent force perturbations. For control purposes, aerodynamic compensation can be achieved either by direct feedback linearization using measured $T/W\approx3.4$ 6 or by gain-scheduling strategies that modulate control gains under high wind disturbances (Hui et al., 2022). Simulation results demonstrate up to 71% reduction in peak CoM tracking errors in wind when using feedback-linearizing controllers.

3. Thrust Estimation and Observer Design

Accurate online estimation of the turbojets' thrust is critical for flight stability and for avoiding engine saturation. iRonCub-Mk3 employs a grey-box state-space model for the turbojet shaft dynamics,

$T/W\approx3.4$ 7

where $T/W\approx3.4$ 8 is shaft speed, $T/W\approx3.4$ 9 is fuel valve command (proportional to mass-flow), and $m\approx30$ 0 models aerodynamic back-torque (Momin et al., 2022). Thrust output is parameterized as $m\approx30$ 1.

A nine-parameter vector $m\approx30$ 2 is identified offline using high-rate rotational and thrust measurements on a custom test bench, via nonlinear least squares with Tikhonov regularization and Levenberg–Marquardt optimization. Once identified, an Extended Kalman Filter (EKF) is deployed at 1 kHz on the flight controller, ingesting only reactor angular speed to return per-turbine thrust estimates.

Bench validation shows EKF thrust error (mean absolute) within $m\approx30$ 3 of $m\approx30$ 4 on step and ramp profiles, and the method is robust to engine failures (e.g., error $m\approx30$ 5 N under compressor stall). Within the iRonCub-Mk3 control stack, the thrust observer output is used by both a model-predictive thrust allocation algorithm and stability-margin supervisor, directly informing high-level flight control and ensuring at least 20% reserve on engine capacity during maneuvers. EKF-based thrust estimation reduces altitude-tracking RMS error by 15% compared to open-loop estimates (Momin et al., 2022).

4. Control Architecture: MPC, Estimation, and Sensing

A key architectural element is the unified multi-rate Model Predictive Control (MPC) scheme, which linearly parameterizes the centroidal momentum equations and explicitly embeds the slow, nonlinear jet actuator dynamics (Gorbani et al., 22 May 2025, Gorbani et al., 1 Jun 2025). The controller state contains CoM position, momenta, base orientation, jet thrusts and rates, and offset integrators for bias-free tracking. The principal design features are:

Actuation bandwidth split: Joint position references at 1,000 Hz; turbojet throttle commands at 10 Hz (Gorbani et al., 1 Jun 2025).
State feedback: Incorporates IMU (Xsens MTI-670G), vision-based position/velocity (Intel RealSense T265), and F/T sensor array (on forearms and jetpack). Fusion via UKF at up to 200 Hz for base pose and 10 Hz for thrust.
Thrust dynamics: Embedded as second-order nonlinear ODEs in the prediction model, essential for realistic jet response. Control input is separated into fast (joints) and slow (jet) channels.
Prediction and optimization: MPC horizon employs variable-knot discretization (N=17 over 1 s), with online QP solves via OSQP achieving average cycle time 2.18 ms (max 4.45 ms) on standard CPUs (Gorbani et al., 22 May 2025).
Cost function: Penalizes CoM tracking error, centroidal momentum error, attitude error, control effort, and bias integrals for offset-free tracking; constraints include jet and joint limits.

In simulation, the unified MPC demonstrates robust disturbance rejection (e.g., 300 Nm torque + 50 N impulse), minimum-jerk trajectory following (CoM MAE $m\approx30$ 60.15 m in $m\approx30$ 7), and ablation studies confirm that including explicit jet dynamics is necessary for stability. Single-rate MPC (no multi-rate splitting) increases position error by ≈30% and amplifies attitude oscillations by an order of magnitude (Gorbani et al., 22 May 2025).

5. CAD-Driven Co-Design and Structural Optimization

The mechanical design of iRonCub-Mk3 leverages a systematic CAD-based co-design approach to jointly optimize geometry, inertial parameters, structural safety, and flight control performance (Vanteddu et al., 2024, Vanteddu et al., 18 Sep 2025). The primary pipeline components are:

Geometric parameterization: Key link subsystems (jetpack brackets, forearm supports) are specified by up to eight integer geometry variables (e.g., jet angle, offset, limb extension; see Table below).
Objective functions: Minimize centroidal momentum error $m\approx30$ 8, joint-velocity error $m\approx30$ 9, and time-averaged total thrust $\approx$ 0 across flight scenarios, subject to minimum FEM safety factors (SF ≥ 10) and QP control feasibility.
Structural validation: Automated FEM on merged STL meshes quantifies von Mises stress under 250 N jet load (per turbine), with instantaneous rejection of geometries failing SF filtering.
CAD–URDF integration: Each geometry yields a fully updated URDF (collision/visual meshes, inertial tensors) for batch simulation in Gazebo or MuJoCo.

Sample results for four Pareto-optimal geometries (from (Vanteddu et al., 2024)):

Variant	Angle (°)	Distance (mm)	Offset (mm)	Length (mm)	$\approx$ 1 Reduction	$\approx$ 2 Reduction
Original	15	42	80	108	—	—
Optim1	1	47	88	50	Yes	Yes
Optim3	1	48	100	130	Best trade-off	Best trade-off

Inertial parameter variations remain within ±10% mass and 15% principal moments. Across the 75 Pareto-optimal candidates, all achieve stress below 46.2 MPa (SF ≥ 10), and flight simulations confirm up to 20% improvement in momentum tracking and 15% reduction in thrust consumption versus baseline Mk3.

Additionally, a large-scale Design of Experiments on 5,000 CAD models, followed by k-means centroid clustering, enables multi-objective tuning of co-design parameters and control weights via NSGA-II (Vanteddu et al., 18 Sep 2025). On the 100-centroid front, dominant trade-offs are observed: variants achieve up to 80% reduction in tracking error (at +50% energy cost), or 50% energy savings (at elevated error), with several “knee” points halving both error and energy relative to the original design.

6. Experimental Results and System-Level Insights

Simulation and first physical takeoff tests validate the iRonCub-Mk3 platform and control approaches. Experimentally, the robot is crane-suspended on a rooftop, with tethers and heat-exclusion safety protocols. Recorded telemetric data confirms:

Lift-off and trajectory tracking: Peak vertical error ≈0.15 m, lateral drift ≤0.6 m, orientation error $\approx$ 38° (Gorbani et al., 1 Jun 2025).
Robustness to unmodeled effects: The multi-rate MPC tolerates significant unmodeled jet lag, wind gusts (1 m/s), and structural vibration, though IMU/F/T sensor accuracy is degraded by turbine-induced vibration.
Control margins: Stability margin throughout flight is maintained (linearized MPC Hessian singular values in [0.1, 10]), and α parameter ramp throughout take-off ensures smooth transition from ground contact to aerial phase.
EKF/UKF-based thrust and pose estimation: Enables closed-loop flight under substantial process and sensor noise.

Key challenges encountered include the complex nonlinear and time-varying jet dynamics (requiring possible future online parameter adaptation in the UKF), vibration-induced estimation errors, thermal effects on sensors, and sim-to-real disparities arising from unmodeled aerodynamics and ground effects.

7. Future Directions and Open Challenges

Ongoing work on iRonCub-Mk3 and derived platforms focuses on several fronts (Gorbani et al., 1 Jun 2025, Gorbani et al., 22 May 2025):

Advanced estimation: Online adaptive identification for thrust models within Kalman filters to counteract parameter drift under high-temperature/long-duration operations.
Sensor robustness: Vibration isolation for inertial sensors and leverage of vision-based drift correction to mitigate mechanical/thermal noise.
Robust control: Extension of MPC to robust/stochastic formulations, explicit disturbance modeling, and adaptive gain scheduling based on real-time model uncertainty.
Morphological expansion: Integration of articulated hands/forearms for manipulation-in-flight scenarios, and material transitions (e.g., carbon fiber) for weight reduction and endurance.
Full untethered flight: Progression toward untethered operation with onboard fuel and greater system autonomy once structural/material and estimation bottlenecks are resolved.

A plausible implication is that the co-design, MPC, and observer frameworks developed for iRonCub-Mk3 are broadly transferable to future jet-powered or high-thrust humanoids, providing scalable methods for structural–control integration under stringent flight constraints.