InnoCube 3U Nanosatellite Platform

• InnoCube 3U Nanosatellite is a modular CubeSat featuring a 3U form factor that integrates quantum key distribution, advanced battery and GNSS technologies, and robust COTS components.
• The platform demonstrates a cutting-edge combination of wireless intra-satellite data bus, high-precision orbit determination, and both classical PD and DRL-based attitude control systems.
• InnoCube validates a cost-effective, modular approach for nanosatellite missions, enhancing secure quantum communications and autonomous control in low Earth orbit.

The InnoCube 3U Nanosatellite is a CubeSat-class nanosatellite platform designed for advanced on-orbit technology demonstration missions, with substantial contributions to both quantum communication and autonomous satellite control research. It features a standard 3U (100×100×340 mm) mechanical form factor, subsystem modularity, and mission architectures integrating commercial off-the-shelf (COTS) components. InnoCube has served both as a platform for a fully wireless SKITH data bus, advanced battery technology demonstration, precise orbit determination with software-defined GNSS, and as a testbed for quantum key distribution (QKD) uplink and AI-driven attitude control in low Earth orbit (Neumann et al., 2017, Djebko et al., 22 Dec 2025).

1. Mission Objectives and Configuration

The InnoCube 3U CubeSat was developed as a joint effort by Julius-Maximilians-Universität Würzburg and Technische Universität Berlin, launched on January 14, 2025 to a 520 km sun-synchronous orbit (Djebko et al., 22 Dec 2025). Its mission profile included:

• Primary Objectives:
  • Demonstrate the SKITH fully wireless intra-satellite data bus.
  • Test “Wall#E”, a solid-state battery embedded within a fiber-reinforced spacecraft wall.
  • Operate the EPISODE software-defined GNSS receiver for high-precision orbit determination.
• Secondary Payloads: Amateur radio transponder and diverse experimental modules.
• ADCS Mission Goals: Implement inertial pointing (±90° anti-nadir GNSS antenna), laser-ranging tracking (±17° accuracy), and technology demonstrations for satellite attitude control.

Physical specifications include a 3U form factor (100×100×340 mm), ≈4.2 kg total mass, four deployable solar panels, 12 Ah battery capacity, and UHF communications at 433 MHz. The on-board computer (Silicon Labs EFR32FG12 ARM® Cortex-M4 @40 MHz, 1 MB flash/256 kB RAM) runs RODOS RTOS, with all modules interfaced via SKITH (Djebko et al., 22 Dec 2025).

2. Mechanical Structure and System Layout

The mechanical and subsystem integration follows a strict volumetric partition (see table below), supporting optical payloads for QKD as well as the ADCS, RF, and power subsystems (Neumann et al., 2017).

Subsystem	Volume (U)	Mass (g)
Optics + Detectors	1.75	750
Measurement Control	0.37	425
ADCS + Beacon + CPU/GPS	0.7	1,100
RF (S-band + UHF)	0.32	215
Battery + Frame	0.1	900
Margin / Harness	0.08	369
Total	3.25	3,759

The 10 cm Cassegrain telescope is mounted on the Z+ end-face for QKD, while panels and radiators are arranged for optimal APD thermal regulation and energy management. COTS component selection with space heritage governs all mission-critical subsystems, enforcing a balance between mass, volume, and structural robustness (Neumann et al., 2017).

3. Quantum Communication Payload and Link Budget

InnoCube was designed as a fully-integrated QKD nanosatellite enabling ground-to-space quantum key exchange using an uplink configuration. The payload comprises a 10 cm Cassegrain receiver telescope, phase shifter, dichroic and PBS optics, Si-APDs (PDM020, ηB ≈ 15%, tB ≈ 37 ps, cooled to –30°C), a liquid-crystal half-wave plate (LC-HWP) for basis switching (tSB ≈ 100 µs), and supporting measurement-control electronics (Neumann et al., 2017).

The system implements both entanglement-based (E91) and decoy-state (DSP) QKD protocols, with secure key rate calculations based on Ma-Lo security formulae. For a 500 km, 30° inclination orbit, the up-link budget combines diffraction, atmospheric turbulence (median Fried parameter r₀ = 20 cm), and plane-parallel extinction (β ≈ 0.22 at 810 nm). At zenith, uplink losses are ΛL(φ=0) ≈ –35 dB, increasing to –43 dB at φ = 60°.

Per-pass secure key generation rates are numerically:

• E91: 1–5 kbps, 10–40 kbit sifted keys per pass, ~4 Mbit/year cumulative.
• DSP: 5–20 kbps, 30–140 kbit per pass, ~13 Mbit/year cumulative.

The link budget is explicitly formulated as:

$$Λ_{tot} = η_A Λ_H^2 Λ_{TA} Λ_L Λ_{PB} Λ_{TB} Λ_{OB} Λ_{SB} Λ_{SYN} η_B$$

where each factor accounts for quantum efficiency, optical throughput, pointing-induced losses (ΛPB ≈ –2.5 dB @σB ≈ 40 µrad), telescope obscuration, and protocol-specific parameters (Neumann et al., 2017).

4. Attitude Determination and Control Subsystem (ADCS)

Both classical and deep learning-based ADCS architectures have been implemented and demonstrated.

Classical PD Approach

The classical attitude controller uses a quaternion-based error metric:

$$\delta q = q_t^{-1} \otimes q = [\delta q_s; \delta q_v]$$

with the control law for reaction wheel torques:

$$\tau_{rw} = -k_p \operatorname{sign}(\delta q_s) \delta q_v - k_d (1 + \|\delta q_v\|^2)\omega$$

where $k_p=[670, 670, 670\cdot0.2]$, $k_d=[2500, 2500, 2500\cdot0.2]$; $\omega$ is the body rate. This controller is manually tuned to meet mission pointing tolerances (inertial: ±90°, laser: ±17°). Hardware comprises 3 orthogonal reaction wheels (2 mNm max torque), ferrite-core magnetorquers (±0.35 A·m²), ST ASM330LHH MEMS gyros, and a PNI RM3100 magnetometer. Typical performance achieves sub-degree precision for standard maneuvers (Djebko et al., 22 Dec 2025, Neumann et al., 2017).

AI-Based DRL Controller

The LeLaR agent represents the first in-orbit demonstration of a satellite attitude controller trained via deep reinforcement learning (DRL). The controller is trained entirely in simulation (dynamics at 10 Hz, control loop at 1 Hz), employing domain randomization (inertia, orbit, sensor noise), and reward shaping based on quaternion error, rate penalty, and action smoothing. The agent uses PPO-derivative “SkipPPO” training, with the deployed model requiring ~105 kB (inference ~38 ms) and running on the Cortex-M4 OBC.

A safety cage monitors angular rate, wheel speeds, and torque; violation disables controller and wheels. Adaptive to on-orbit discrepancies (e.g., wheel speed estimation errors, RW dead time), the agent maintained robust performance under variable hardware conditions (Djebko et al., 22 Dec 2025).

5. In-Orbit Demonstration and Comparative Results

The in-orbit performance of both PD and LeLaR controllers is quantitatively characterized:

• LeLaR (flight-agent) steady-state errors:
  • Random initial → target (1,0,0,0): settled in 114 s, steady-state yaw (σ=0.11°), pitch (σ=0.08°), roll (σ=0.18°) over 276 s.
  • Following 20 s wheel dead time: settled in 58 s, steady-state errors Yaw (σ=0.09°), Pitch (σ=0.07°), Roll (σ=0.07°) over 69 s.
  • Repeated identical maneuvers (7×): all settled <1° in 44–114 s; PD controller only 3/7 <1°, roll σ up to 0.92°.
  • Six diverse-attitude maneuvers: all <1°, σ_yaw 0.05–0.15°, σ_pitch 0.04–0.25°, σ_roll 0.08–0.29°; PD: 2/6 within <1°, errors up to ±2°.

LeLaR outperformed PD in both settling time, steady-state error, and resilience to actuator/sensor anomalies without post-launch retuning. The deployment validates DRL-based control as a robust, adaptable approach under real-world orbital perturbations (Djebko et al., 22 Dec 2025).

6. Design Trade-Offs and Engineering Lessons

Subsystem and component choices on InnoCube were motivated by a set of quantifiable trade-offs:

Component	Option A (Chosen)	Option B	Impact
Detector	Si-APD (15%, 37 ps)	SNSPD (70%, 15 ps)	APD: no cryo, simple; SNSPD: high SNR, heavy, power
Basis switch	LC-HWP (100 µs)	Pockels cell (1 µs)	LC-HWP: low power/mass; PC: high voltage, losses
ADCS	COTS star-tracker + reaction wheels	Custom fast steer	COTS: proven, low risk; custom: better jitter, cost

Further distilled design principles include: maximizing simplicity by centralizing quantum complexity in the OGS for a generic satellite; minimizing detector area for noise; balancing APD quantum efficiency and timing jitter for SNR optimization; prioritizing COTS for risk/cost reduction; transparent link budget expression for modular analysis; strict SWaP (size, weight, power) discipline through operations scheduling; and modular architecture as a reusable CubeSat Q.Com template (Neumann et al., 2017).

A plausible implication is that InnoCube’s modular architecture enables cost-effective scaling for QKD constellations, and its DRL ADCS pipeline suggests a generalizable approach for autonomous control in future CubeSat missions.

7. Future Directions and Significance for Nanosatellite Research

The InnoCube platform evidences the feasibility of deploying fully functional, mission-adaptable nanosatellites with SWaP-constrained quantum and AI-driven subsystems. Key documented future trajectories encompass:

• Extension of experimental windows (e.g., >15 min) for full-scale momentum management and advanced control modes, such as spinning-sun pointing or formation flying.
• Refinement of simulation environments using in-orbit telemetry to enhance Sim2Real transfer fidelity for DRL-based control.
• Automation of the DRL training and deployment pipeline to allow rapid, drop-in controller generation across CubeSat variants.
• Expansion of QKD satellite constellations leveraging the cost-effective, sub-US\$1M platform model and modular design assets (Djebko et al., 22 Dec 2025, Neumann et al., 2017).

By formalizing transparent, modular, and COTS-based Q.Com and control architectures, InnoCube provides a validated template for future research and commercial deployments in secure global quantum communication and autonomous nanosatellite operation.