Trustworthy AI Control Framework Based on Physics-Informed Neural Networks (PINNs)
PINN-trained controllers with stability guarantees, generated without any training data; contraction metrics on Riemannian manifolds (RCCM) for singularity-free robust control. Accepted at CDC 2026 and AIAA SciTech 2027.
Period: 2025/08 – 2026/03 · University of Toronto Institute for Aerospace Studies
- Led a two-person team (1 Ph.D. student, 1 master’s student) and proposed a nonlinear control training method based on physics-informed neural networks (PINNs) that enables neural networks to automatically generate robust control algorithms from a system’s multibody dynamics equations without any training dataset, enabling rapid generation of control schemes for complex systems. Compared with conventional machine-learning control, the PINN-generated controllers strictly satisfy the theoretical requirements of control stability, yielding reliable neural-network-based control.
- Through differential geometry, combined PINNs with Riemannian manifold theory to achieve singularity-free control of the spatial attitude of drones and robots. Combined with contraction analysis control theory, transformed the robust control problem into the solution of a partial differential equation (PDE), and then solved the PDE with PINNs to obtain control networks with stability guarantees.
- Proposed, for the first time, a robust control theory combining PINNs with Riemannian manifolds (Robust Control Contraction Metrics, RCCM), which still guarantees control accuracy in the presence of unknown external disturbances and parameter uncertainties.
- The PINN-based AI control framework proposed in this research can be widely applied to the control of unmanned systems and embodied intelligence.
- Results have been accepted by the top international control conference, the IEEE Conference on Decision and Control (CDC 2026), and by the AIAA 2027 SciTech Forum: “Neural Robust Control on Lie Groups Using Contraction Methods” (CDC 2026) and “Contraction-based Neural Control for Cooperative Aerial Payload Transportation with Variable-length Cables” (AIAA SciTech 2027).
Theory at a glance
Both papers turn stability into conditions on a Riemannian metric, then let neural networks find a metric and a controller that satisfy those conditions at sampled states. No demonstration data is needed, only the system’s equations of motion.
Background
A. Contraction analysis: what it is. Lyapunov analysis asks whether trajectories converge to one equilibrium. Contraction analysis (Lohmiller and Slotine, 1998) asks a stronger question: do any two neighbouring trajectories converge to each other? It looks at the differential (virtual) dynamics \(\delta\dot{x} = A(x)\,\delta x\) with \(A = \partial f / \partial x\), and measures the length of \(\delta x\) with a Riemannian metric \(M(x) \succ 0\). If that length shrinks at a uniform rate \(\lambda\),
\[\frac{d}{dt}\big(\delta x^\top M(x)\,\delta x\big) \le -2\lambda\, \delta x^\top M(x)\,\delta x \quad \Longleftrightarrow \quad \dot{M} + \langle M A \rangle + 2\lambda M \preceq 0 ,\]then all trajectories converge to each other exponentially. With \(\underline{m} I \preceq M \preceq \overline{m} I\),
\[\lVert x_1(t) - x_2(t) \rVert \le \sqrt{\overline{m}/\underline{m}}\; e^{-\lambda t}\, \lVert x_1(0) - x_2(0) \rVert .\]Why use it for control.
- Tracking, not just stabilization. If all trajectories converge to each other, the system converges to whichever feasible reference \((x^*, u^*)\) the planner sends. There is no need to build a new Lyapunov function for every trajectory.
- Pointwise conditions. The certificate is a matrix inequality at each state \(x\). It can be checked, or penalized in a loss, one sampled state at a time, which is exactly what neural-network training needs.
- Robustness with numbers. Disturbances enter the differential dynamics as an extra input, which gives explicit gain bounds and error tubes rather than a qualitative “it is stable”.
B. Control contraction metrics (CCM). Manchester and Slotine (2017) extended contraction to systems with inputs, \(\dot{x} = f(x) + B(x)u\). A CCM is a metric for which some differential feedback \(\delta u = K(x)\,\delta x\) makes the closed loop contract, \(\dot{M} + \langle M(A + BK) \rangle + 2\lambda M \preceq 0\). With the dual metric \(W = M^{-1}\) and \(B_\perp\) spanning the directions the inputs cannot reach (\(B_\perp^\top B = 0\)), sufficient conditions that no longer depend on \(u\) are
\[B_\perp^\top \Big( -\partial_f W + \big\langle \tfrac{\partial f}{\partial x} W \big\rangle + 2\lambda W \Big) B_\perp \prec 0, \qquad \partial_{b_i} W - \big\langle \tfrac{\partial b_i}{\partial x} W \big\rangle = 0 .\]The first says the system must contract on its own in the directions the controller cannot push. The second says the metric does not change along the input directions (the \(b_i\) are Killing fields), so the controller can make those directions contract. If such a metric exists, a controller that tracks any feasible trajectory exists too. The robust version (RCCM; Manchester and Slotine, 2018) adds the disturbance input and turns contraction into the \(\mathcal{L}_\infty\) gain bound used below.
C. Lie groups. A Lie group is a smooth manifold that is also a group. The attitude of a drone, a payload or a robot link is a rotation,
\[SO(3) = \{\, R \in \mathbb{R}^{3\times 3} : R^\top R = I_3,\ \det R = 1 \,\}, \qquad \dot{R} = R\, \omega^\wedge ,\]where \(\omega^\wedge\) is the skew-symmetric matrix of the angular velocity \(\omega\). Coordinates such as Euler angles break down: they are singular at \(\pm 90^\circ\) pitch (gimbal lock) and distort distances, and quaternions cover each rotation twice. Working on the group itself is globally valid. Every Lie group is parallelizable: the velocity at any \(R\) is \(R\,\xi^\wedge\) for some \(\xi \in \mathbb{R}^3\), so a single smooth basis \(S(x)\) spans the tangent space everywhere, and every tangent vector can be written as \(\delta x = S(x)\,v\) with intrinsic coordinates \(v\).
D. Why CCM needs the Lie-group view. CCM theory is written for \(\mathbb{R}^n\). A rotation stored as the 9 entries of \(R\) has only 3 degrees of freedom, so a Euclidean metric on \(\mathbb{R}^9\) would also have to contract in 6 directions that leave \(SO(3)\) and have no physical meaning. The papers below instead write the differential dynamics in the intrinsic coordinates \(v\) and search for a smaller metric \(\mathcal{M}(x)\) on the tangent space only, lifted back with the projection \(P_S = S(S^\top S)^{-1}\):
\[\delta x = S\,v, \qquad B = S\,E, \qquad B_w = S\,E_w, \qquad M(x) = P_S(x)\, \mathcal{M}(x)\, P_S(x)^\top .\]As a result, the certificate is valid for every attitude with no singularities. Errors are measured as distances on the manifold, using the Lie-algebra error \(\varepsilon_r = \tfrac{1}{2}\big(R^{*\top} R - R^\top R^*\big)^\vee\) for rotations. The matrix conditions also shrink: for the three-drone payload the state is a 24-dimensional manifold embedded in \(\mathbb{R}^{30}\), so the networks learn \(24\times 24\) metrics instead of \(30\times 30\) ones.
Results
1. Systems on Lie groups. The state lives on a manifold \(\mathcal{X}\) (for a quadrotor, \(\mathbb{R}^3 \times SO(3)\) and its velocities), with disturbance \(w\), output \(z\) and a feedback law around a reference \((x^*, u^*)\):
\[\dot{x} = f(x) + B(x)\,u + B_w(x)\,w, \qquad z = g(x,u), \qquad u = k(x,x^*) + u^* .\]As in D, differential lengths are measured on the tangent space, \(V = v^\top \mathcal{M}(x)\, v\) with \(\delta x = S(x)\,v\).
2. Robust control contraction metric (RCCM). The closed loop has an \(\mathcal{L}_\infty\) gain \(\alpha\) from disturbance to output if there are a uniformly bounded metric \(\mathcal{M}(x)\) and constants \(\lambda, \mu > 0\) such that
\[R_1 = \begin{bmatrix} \dot{\mathcal{M}} + \langle \mathcal{M}\mathcal{A} \rangle + 2\lambda\mathcal{M} & \mathcal{M}E_w \\ E_w^\top \mathcal{M} & -\mu I_p \end{bmatrix} \preceq 0, \qquad R_2 = \begin{bmatrix} 2\lambda\mathcal{M} - \alpha^{-1}\mathcal{C}^\top\mathcal{C} & 0 \\ 0 & (\alpha-\mu) I_p \end{bmatrix} \succeq 0,\]where \(\mathcal{A}\) and \(\mathcal{C}\) are the closed-loop differential dynamics and output map in intrinsic coordinates.
3. What it guarantees. The tracking error is bounded by the size of the disturbance,
\[\lVert z - z^* \rVert^2_{\mathcal{L}_\infty^{[0,T]}} \le \alpha^2 \, \lVert w - w^* \rVert^2_{\mathcal{L}_\infty^{[0,T]}} + \beta\big(x(0), x^*(0)\big),\]so after transients the output stays inside a tube \(\Omega(z^*) = \{\, y : \lVert y - z^* \rVert \le \alpha \bar{w} \,\}\) around the reference.
4. Learning the certificate. The controller and the dual metric \(\mathcal{W} = \mathcal{M}^{-1}\) are neural networks, built so that \(k_{nn}(x,x) = 0\) and \(\mathcal{W} \succeq \overline{m}^{-1} I\) hold by construction:
\[k_{nn}(x,x^*) = K_1(x,x^*;\theta_{k_1}) \tanh\!\big(K_2(x,x^*;\theta_{k_2})\, \varepsilon(x,x^*)\big), \qquad \mathcal{W}(x) = \Theta(x;\theta_w)^\top \Theta(x;\theta_w) + \overline{m}^{-1} I .\]They are trained jointly on states sampled from the operating region, with a loss that penalizes every violation of the conditions:
\[\mathcal{L}_k = \mathrm{L_{PD}}(-R_1) + \mathrm{L_{PD}}(R_2) + \mathrm{L_{PD}}(-C_1) + \sum_i \lVert C_{2,i} \rVert_F + \sum_j \lVert C_{3,j} \rVert_F + \mathrm{L_{PD}}(\underline{m}^{-1} I - \mathcal{W}) + \mathrm{ReLU}(\alpha - \underline{\alpha}),\]where \(\mathrm{L_{PD}}\) penalizes a matrix that is not positive semidefinite, \(C_1\) asks the uncontrolled system to contract in directions the inputs cannot reach, and \(C_{2,i}, C_{3,j}\) make the input and disturbance directions Killing fields of the metric (CDC 2026).
5. Application: cooperative payload transport with variable-length cables (AIAA SciTech 2027). For three drones carrying a rigid-body payload, the equations of motion are arranged so that the payload and the cable lengths are driven by separate control channels. The payload subsystem \(\Sigma_p\) (state in \(\mathbb{R}^{21} \times SO(3)\)) gets a neural CCM controller that enforces
\[C_{CCM} = \dot{\mathcal{M}} + \langle \mathcal{M}\mathcal{A} \rangle + 2\lambda\mathcal{M} \preceq 0,\]while the cable-length subsystem is a simple double integrator with its own tracker,
\[\Sigma_l: \quad \frac{d}{dt}\begin{bmatrix} l \\ \dot{l} \end{bmatrix} = \begin{bmatrix} \dot{l} \\ 0 \end{bmatrix} + \begin{bmatrix} 0 \\ I_N \end{bmatrix} \tilde{f}_{\parallel},\]so the formation can shorten its cables to fly through height-constrained gates, as in the simulation above.