Guidance-and-Control CNet

• Guidance-and-Control CNet is a family of neural architectures that approximates optimal state-feedback laws, directly mapping sensor and state inputs to control commands.
• They leverage optimal control principles and are trained via imitation learning and reinforcement learning to achieve high performance in diverse applications from interplanetary transfers to quadcopter racing.
• Their design emphasizes onboard computational efficiency and robust adaptation to uncertainties, enabling real-time control in resource-constrained environments.

Guidance-and-Control CNet denotes a class of neural guidance and control architectures in which sensor or state information is mapped directly to control commands according to an underlying optimality principle. In the literature, the term appears as G&CNet, G&CNET, GCNET, and integrated guidance-and-control CNet, but the shared idea is consistent: the network approximates an optimal state-feedback law and shifts onboard from tracking precomputed guidance profiles toward computing control actions in real time. Reported applications span low-thrust interplanetary transfers, pinpoint landing, close-proximity operations, quadcopter racing, and lunar powered descent, with training based either on optimal-control solutions generated offline or on reinforcement learning in simulation (Izzo et al., 2023, Izzo et al., 2020, Gaudet et al., 2021).

1. Optimal-control basis

The canonical formulation treats the vehicle as a dynamical system with state vector $x(t)\in\mathbb{R}^n$ and control vector $u(t)\in\mathbb{R}^m$ governed by

$$\dot x(t) = f(x(t),u(t)), \qquad t_0 \le t \le t_f,$$

and seeks the control history $u(\cdot)$ that minimizes

$$J[u(\cdot)] = \int_{t_0}^{t_f} L(x(t),u(t))\,dt + \Phi(x(t_f)),$$

subject to boundary conditions $x(t_0)=x_0$, $x(t_f)\in\mathbb{X}_f$, and possibly $u(t)\in\mathbb{U}$, $x(t)\in\mathbb{X}$. The cited works instantiate this template with time-optimal objectives $L(x,u)=1$, fuel- or mass-optimal objectives such as $u(t)\in\mathbb{R}^m$0, and energy-optimal objectives such as $u(t)\in\mathbb{R}^m$1 (Izzo et al., 2023).

For deterministic problems, Pontryagin’s Minimum Principle yields the Hamiltonian

$u(t)\in\mathbb{R}^m$2

with co-state dynamics $u(t)\in\mathbb{R}^m$3 and pointwise optimality condition

$u(t)\in\mathbb{R}^m$4

In stochastic settings, the value function satisfies the Hamilton–Jacobi–Bellman equation

$u(t)\in\mathbb{R}^m$5

A guidance-and-control CNet is then trained so that the neural feedback law $u(t)\in\mathbb{R}^m$6 approximates $u(t)\in\mathbb{R}^m$7 through

$u(t)\in\mathbb{R}^m$8

This encapsulates the solution of a broad class of optimal control problems into a single neural feedback law (Izzo et al., 2023).

Two concrete optimal-control structures recur in the literature. In mass-optimal low-thrust transfer, the switching function

$u(t)\in\mathbb{R}^m$9

induces a “bang–off–bang” law for the thrust magnitude, while the thrust direction is

$$\dot x(t) = f(x(t),u(t)), \qquad t_0 \le t \le t_f,$$0

In pure time-optimal quadcopter flight, the Hamiltonian is affine in the control, and Pontryagin’s condition places the minimizer at the box constraints $$\dot x(t) = f(x(t),u(t)), \qquad t_0 \le t \le t_f,$$1, producing “bang–bang” control (Izzo et al., 2020, Origer et al., 2023).

2. Network formulations and signal pathways

The most general G&CNet description in the cited corpus is a feed-forward multilayer perceptron whose inputs may include the state vector $$\dot x(t) = f(x(t),u(t)), \qquad t_0 \le t \le t_f,$$2, task or environment parameters $$\dot x(t) = f(x(t),u(t)), \qquad t_0 \le t \le t_f,$$3, and optionally online estimates of unmodeled disturbances $$\dot x(t) = f(x(t),u(t)), \qquad t_0 \le t \le t_f,$$4, and whose outputs are direct control commands $$\dot x(t) = f(x(t),u(t)), \qquad t_0 \le t \le t_f,$$5 such as thrust-vector components, body-rates, or motor speeds. A typical “small” G&CNet has input dimension $$\dot x(t) = f(x(t),u(t)), \qquad t_0 \le t \le t_f,$$6 with often $$\dot x(t) = f(x(t),u(t)), \qquad t_0 \le t \le t_f,$$7–$$\dot x(t) = f(x(t),u(t)), \qquad t_0 \le t \le t_f,$$8 total inputs, three fully connected layers of approximately $$\dot x(t) = f(x(t),u(t)), \qquad t_0 \le t \le t_f,$$9 neurons each, hidden activations chosen from ReLU or tanh, and a linear output layer of dimension $u(\cdot)$0 (Izzo et al., 2023).

Variant	Inputs	Outputs
Generic G&CNet	$u(\cdot)$1, parameters $u(\cdot)$2, optionally $u(\cdot)$3	Direct control commands $u(\cdot)$4
Interplanetary policy network	$u(\cdot)$5	$u(\cdot)$6 and three direction weights
Neural ODE transfer network	$u(\cdot)$7	Unit thrust direction
Neural ODE landing network	$u(\cdot)$8	Throttle and thrust direction
Quadcopter CNet	$u(\cdot)$9, optionally $$J[u(\cdot)] = \int_{t_0}^{t_f} L(x(t),u(t))\,dt + \Phi(x(t_f)),$$0 or look-ahead waypoint data	Four actuator inputs
Lunar powered-descent CNet	18-dimensional observation including $$J[u(\cdot)] = \int_{t_0}^{t_f} L(x(t),u(t))\,dt + \Phi(x(t_f)),$$1, $$J[u(\cdot)] = \int_{t_0}^{t_f} L(x(t),u(t))\,dt + \Phi(x(t_f)),$$2, $$J[u(\cdot)] = \int_{t_0}^{t_f} L(x(t),u(t))\,dt + \Phi(x(t_f)),$$3, quaternion terms, and $$J[u(\cdot)] = \int_{t_0}^{t_f} L(x(t),u(t))\,dt + \Phi(x(t_f)),$$4	Four-engine thrust commands
Memristor GCNET	$$J[u(\cdot)] = \int_{t_0}^{t_f} L(x(t),u(t))\,dt + \Phi(x(t_f)),$$5	$$J[u(\cdot)] = \int_{t_0}^{t_f} L(x(t),u(t))\,dt + \Phi(x(t_f)),$$6

Within that common template, individual papers make sharply different architectural choices. The interplanetary transfer work uses softplus activations and trains both a policy network and a value-function network, with the policy network outputting throttle and thrust-direction weights (Izzo et al., 2020). The Neural ODE refinement work uses three fully connected hidden layers with $$J[u(\cdot)] = \int_{t_0}^{t_f} L(x(t),u(t))\,dt + \Phi(x(t_f)),$$7 neurons each, softplus in all hidden layers, linear outputs for transfer direction, and a mixed linear-plus-sigmoid output for landing so that $$J[u(\cdot)] = \int_{t_0}^{t_f} L(x(t),u(t))\,dt + \Phi(x(t_f)),$$8 is enforced directly (Origer et al., 2024). The quadcopter racing architecture uses three $$J[u(\cdot)] = \int_{t_0}^{t_f} L(x(t),u(t))\,dt + \Phi(x(t_f)),$$9-unit ReLU layers followed by a four-sigmoid output, and augments the input either with an estimated rotor-speed limit or with a two-waypoint look-ahead vector (Origer et al., 2023). The lunar powered-descent CNet departs more substantially from the feed-forward pattern by inserting a gated recurrent unit between fully connected layers, with online mean-variance normalization of all observation components (Gaudet et al., 2021). The memristor study returns to a compact fully connected topology with three $x(t_0)=x_0$0-neuron hidden layers, softplus hidden activations, and tanh output (Rudge et al., 2 Sep 2025).

This suggests that “Guidance-and-Control CNet” denotes a family of architectures organized by control objective and onboard constraints rather than a single fixed network design.

3. Training regimes and data generation

The most established training paradigm is imitation learning from optimal trajectories. The standard workflow is to solve the deterministic optimal-control problem by indirect methods via Pontryagin or by direct transcription, record state–action pairs, and minimize a supervised loss such as mean-squared error or cosine-similarity loss. Reported dataset scales range from tens of thousands of trajectories and approximately $x(t_0)=x_0$1–$x(t_0)=x_0$2 samples to substantially larger corpora produced by Backward Generation of Optimal Examples (BGOE), which can amplify a single nominal solution into $x(t_0)=x_0$3–$x(t_0)=x_0$4 perturbed trajectories at comparable computational cost and, in the interplanetary study, produce datasets of size up to $x(t_0)=x_0$5 in a few hours instead of years (Izzo et al., 2023, Izzo et al., 2020).

Loss design depends on the control parameterization. For thrust-direction learning, the literature frequently uses a cosine-similarity loss,

$x(t_0)=x_0$6

while fuel-optimal scenarios with both throttle and direction add an MSE term on throttle. Data augmentation with DaGGER is proposed to reduce deviation from nominal paths, and the comparative study emphasizes the use of multiple BGOE “bundles” with varying costate perturbation magnitudes and times of flight to mitigate distribution shift (Izzo et al., 2023, Holt et al., 22 Jul 2025).

A second training paradigm formulates the problem as a Markov Decision Process and learns a stochastic policy with policy-gradient or actor–critic methods such as PPO. In the comparison study, the policy is Gaussian, training uses PPO with clipped surrogate objective, reward redistribution is introduced to avoid sparse terminal rewards, and continuous integration recomputes the action during numerical integration rather than holding it fixed between RL action times. The same study reports that BC training is faster and more sample efficient, whereas RL can better adapt to stochastic conditions and can discover solutions that improve upon suboptimal expert demonstrations (Holt et al., 22 Jul 2025).

A third training regime refines a behaviourally cloned G&CNet by embedding it in the closed-loop dynamics as a Neural ODE. The network parameters are then updated by minimizing the final-state mismatch

$x(t_0)=x_0$7

using variational equations for $x(t_0)=x_0$8. Reported reductions are substantial: for a nominal interplanetary transfer, position error decreases from approximately $x(t_0)=x_0$9 km to approximately $x(t_f)\in\mathbb{X}_f$0 km and velocity error from approximately $x(t_f)\in\mathbb{X}_f$1 km/s to approximately $x(t_f)\in\mathbb{X}_f$2 km/s; for asteroid landing, position error decreases from approximately $x(t_f)\in\mathbb{X}_f$3 m to approximately $x(t_f)\in\mathbb{X}_f$4 m (Origer et al., 2024).

A related but distinct line of work uses neural networks not as the controller but as a learnable front-end for sequential convex programming. In that formulation, a DNN predicts an initial state-control sequence for the convex solver, reducing the median SCP iteration count from $x(t_f)\in\mathbb{X}_f$5 to $x(t_f)\in\mathbb{X}_f$6 and saving $x(t_f)\in\mathbb{X}_f$7 of computation time in powered landing. The authors explicitly state that, “instead of brutally using the neural networks as the controller,” the network is used to improve the SCP algorithm (Shen et al., 2022). This is not a pure CNet in the end-to-end feedback sense, but it is part of the broader guidance-and-control computational landscape.

4. Demonstrated domains and empirical performance

Reported case studies cover both spacecraft and aggressive aerial robotics, and the numerical claims vary with task, dynamics model, and training paradigm (Izzo et al., 2023, Izzo et al., 2020, Origer et al., 2023, Gaudet et al., 2021, Holt et al., 22 Jul 2025).

Domain	Reported setup	Reported result
Interplanetary low-thrust transfer	Earth–Venus mass-optimal transfer	Spacecraft employing the learned thrust spends only $x(t_f)\in\mathbb{X}_f$8 permil more propellant than the mathematically optimal transfer
Interplanetary low-thrust transfer	BepiColombo-style simulation	Fuel-optimal trajectory tracking error $x(t_f)\in\mathbb{X}_f$9 across test set
Planetary landing	6-DOF Mars-lander analog	Touchdown errors $u(t)\in\mathbb{U}$0 m and $u(t)\in\mathbb{U}$1
Close-proximity operations	Hover-point policy near irregular asteroids	Position error $u(t)\in\mathbb{U}$2 cm
Quadcopter racing	Parrot AR 2.0, 16-DOF model	Lap times within $u(t)\in\mathbb{U}$3 of the true time-optimal solution
Lunar powered descent	5000 test episodes	$u(t)\in\mathbb{U}$4 success rate with mean miss $u(t)\in\mathbb{U}$5 m and mean terminal speed $u(t)\in\mathbb{U}$6 m/s

For low-thrust transfers, the interplanetary G&CNET study reports that the optimal propellant mass can be predicted within an error well within $u(t)\in\mathbb{U}$7, that terminal orbital elements measured by reduced Euclidean distance are $u(t)\in\mathbb{U}$8 in the nominal-start test, and that the success rate is at least $u(t)\in\mathbb{U}$9 for the $x(t)\in\mathbb{X}$0 off-nominal region with $x(t)\in\mathbb{X}$1 (Izzo et al., 2020). The same domain appears again in the general survey, which reports a BGOE-generated dataset of approximately $x(t)\in\mathbb{X}$2 trajectories for an ESA BepiColombo-style problem (Izzo et al., 2023).

For quadcopters, the time-optimal flight study documents a degradation in learnability as the cost approaches pure time-optimality. The control-loss values rise from $x(t)\in\mathbb{X}$3 at $x(t)\in\mathbb{X}$4 to $x(t)\in\mathbb{X}$5 at $x(t)\in\mathbb{X}$6, and the reported mean control error rises from $x(t)\in\mathbb{X}$7 to $x(t)\in\mathbb{X}$8; simulated flight becomes unstable above $x(t)\in\mathbb{X}$9 error when $L(x,u)=1$0. On a $L(x,u)=1$1 m track at $L(x,u)=1$2, the CNet achieves a first lap of $L(x,u)=1$3 s and a steady second lap of $L(x,u)=1$4 s, compared with $L(x,u)=1$5 s and $L(x,u)=1$6 s for the differential-flatness-based minimum snap benchmark controller (Origer et al., 2023).

The BC-versus-RL comparison adds a different empirical perspective. In the continuous-thrust spacecraft tasks, BC more closely replicates expert policy behaviour on deterministic nominal cases, but RL performs better in several stochastic robustness tests. In the fuel-optimal Earth–Mars case, the reported final mass ratio is $L(x,u)=1$7 for BC and $L(x,u)=1$8 for RL against an optimum of $L(x,u)=1$9, with velocity residuals of $u(t)\in\mathbb{R}^m$00 m/s and $u(t)\in\mathbb{R}^m$01 m/s respectively; in the 67P landing case, the RL policy surpasses the indirect local optimum and reduces flight time from $u(t)\in\mathbb{R}^m$02 h to $u(t)\in\mathbb{R}^m$03 h (Holt et al., 22 Jul 2025).

5. Robustness, adaptation, and certification

Robustness in CNet research is addressed at three levels: policy conditioning, training paradigm, and post-training analysis. At the policy level, parametric inputs $u(t)\in\mathbb{R}^m$04 such as gravity uncertainty estimate or actuator limit allow online adaptation to unmodeled effects, and adaptive G&CNets that include $u(t)\in\mathbb{R}^m$05 estimates are reported to remove drift accumulation and maintain stability outside the nominal training manifold. In drone tests, adaptive nets reduced overshoots by more than $u(t)\in\mathbb{R}^m$06 when confronted with an unmodeled motor fault or aerodynamic disturbance. Compared with classical MPC + PID pipelines, G&CNets are reported to achieve similar optimality at $u(t)\in\mathbb{R}^m$07–$u(t)\in\mathbb{R}^m$08 lower onboard computational cost and without requiring online optimization (Izzo et al., 2023).

The BC-versus-RL comparison refines this picture by showing that nominal optimality and stochastic robustness do not coincide. The study’s design guidelines state that BC is fast to train and replicates expert behaviour near-optimally on nominal cases, but requires a high-fidelity expert dataset and has limited generalisation, whereas RL does not require pre-solved experts, is robust to out-of-distribution noise, and can discover novel or improved solutions, at the cost of slower sample efficiency and more difficult reward shaping. It recommends a “BC→RL hybrid” in which BC warm-starts the policy and RL fine-tunes it under stochastic perturbations (Holt et al., 22 Jul 2025).

Certification is treated explicitly in the event-manifold analysis. There, a G&CNET is defined as a neural network trained to approximate the optimal state-feedback policy of a dynamical system, and the central question is whether it satisfies mission-critical terminal requirements under initial-state uncertainty. The proposed method replaces brute-force propagation to a random terminal time with a high-order Taylor map on an event manifold, yielding Event Transition Tensors and a Cauchy–Hadamard convergence radius. Uncertainty propagation is then carried out analytically with moment-generating functions. Reported examples include an interplanetary $u(t)\in\mathbb{R}^m$09 km along the $u(t)\in\mathbb{R}^m$10-axis at the sphere of influence, asteroid-landing radii of convergence of approximately $u(t)\in\mathbb{R}^m$11 kg in mass, $u(t)\in\mathbb{R}^m$12 km in positions, and $u(t)\in\mathbb{R}^m$13 m/s in velocities, and a drone-racing radius that easily covers typical launch dispersions (Origer et al., 2024).

A recurring misconception is that CNet validation is exhausted by Monte Carlo evaluation. The certification work argues the opposite: Monte Carlo can yield the relevant results, but may be insufficient for future certification of neural networks in guidance and control applications. Another misconception is that CNet robustness is synonymous with reinforcement learning. The supervised and adaptive G&CNet studies show that robustness can also be engineered by conditioning on uncertainties, by including disturbance estimates, and by refining cloned policies against closed-loop dynamics (Izzo et al., 2023, Origer et al., 2024, Origer et al., 2024).

6. Onboard deployment, accelerators, and hardware constraints

A central motivation for Guidance-and-Control CNet research is onboard feasibility under tight compute and memory budgets. Reported deployment data include inference at $u(t)\in\mathbb{R}^m$14 Hz on a dual-core $u(t)\in\mathbb{R}^m$15 MHz ARM Cortex-A9 for quadcopter racing, latency $u(t)\in\mathbb{R}^m$16 ms per forward pass on a resource-restricted CPU, and memory footprint $u(t)\in\mathbb{R}^m$17 MB for a $u(t)\in\mathbb{R}^m$18 network. The survey states that such feasibility has been demonstrated on Parrot AR 2.0 drones and is directly transferable to nanosatellites with similar CPUs or FPGAs, explicitly naming OPS-SAT and $u(t)\in\mathbb{R}^m$19-Sat 1. It also points to Loihi and event-cameras as possible neuromorphic substrates for lower-power, continuous-learning G&CNets (Izzo et al., 2023).

The memristor-acceleration work pushes this deployment theme toward in-memory analog computing. Its GCNET is implemented on PCM- or RRAM-based crossbar arrays, with differential conductance encoding, modeled write noise, read noise, conductance drift, and peripheral $u(t)\in\mathbb{R}^m$20-bit DACs and $u(t)\in\mathbb{R}^m$21-bit ADCs. The primary metric is the validation-set loss $u(t)\in\mathbb{R}^m$22. The study reports that moving from $u(t)\in\mathbb{R}^m$23 to $u(t)\in\mathbb{R}^m$24 bit-slices halves the loss from approximately $u(t)\in\mathbb{R}^m$25 to approximately $u(t)\in\mathbb{R}^m$26, that networks can recover nominal accuracy up to approximately $u(t)\in\mathbb{R}^m$27 stuck-at-$u(t)\in\mathbb{R}^m$28 faults after $u(t)\in\mathbb{R}^m$29 epochs of hardware-aware retraining, and that drift is more severe for PCM than for RRAM over $u(t)\in\mathbb{R}^m$30 h and $u(t)\in\mathbb{R}^m$31 h. It also states that the memristor-accelerated GCNET can perform guidance and control inference in $u(t)\in\mathbb{R}^m$32s latency, while present analog non-idealities leave absolute accuracy approximately $u(t)\in\mathbb{R}^m$33 worse than floating-point (Rudge et al., 2 Sep 2025).

This deployment literature broadens the meaning of “Guidance-and-Control CNet.” In its strictest use, the term refers to a neural feedback law that directly outputs control actions. In a broader computational sense, it can also denote architectures in which a neural component embeds offline optimal-control knowledge inside a larger guidance stack, as in the DNN-warm-started SCP system for powered landing (Shen et al., 2022). The overall trajectory of the field suggests a convergence between optimal-control structure, compact neural approximation, and hardware-conscious implementation for real-time autonomy in spacecraft and other resource-constrained robotic platforms.