Neural Contraction Metrics for Robust Estimation and Control: A Convex Optimization Approach
read more
Citations
Review of advanced guidance and control algorithms for space/aerospace vehicles
Safe Control With Learned Certificates: A Survey of Neural Lyapunov, Barrier, and Contraction Methods for Robotics and Control
Robust Controller Design for Stochastic Nonlinear Systems via Convex Optimization
Neural Stochastic Contraction Metrics for Learning-Based Control and Estimation
Learning Certified Control using Contraction Metric
References
Long short-term memory
Convex Optimization
Linear Matrix Inequalities in System and Control Theory
Speech recognition with deep recurrent neural networks
Related Papers (5)
Frequently Asked Questions (13)
Q2. What is the novelty of the NCM approach?
The novelty of the NCM approach lies in that: 1) data points of the optimal contraction metric are sampled offline by solving a convex optimization problem, which minimizes an upper bound of the steady-state Euclidean distance between the perturbed and unperturbed trajectories without assuming any hypothesis function space, and 2) the deep LSTM-RNN is constructed to model the sampled metrics with arbitrary accuracy.
Q3. What is the simplest way to get the limt?
Since the authors have ‖dc(q, t)‖ ≤ b2d and the differential dynamics isδq̇ = (A(x, t)− B1(x, t)B1(x, t)TM(x, t))δq (20) when dc = 0, the authors get limt→∞ ∫ x 0 ‖δq‖ ≤ b2dχ/α by the same proof as for Theorem 3 with (13) replaced by (20), (14) by (18), and Re(t) by Rc(t) = ∫ x 0 ‖ (x, t)δq(t)‖, where M = T .
Q4. What is the result of Lemma 2?
As a result of Lemma 2, the authors can select (x, t) defined in Theorem 1 as the unique Cholesky decomposition of M(x, t) and train the deep LSTM-RNN using only the non-zero entries of the unique upper triangular matrices { (x(ti), ti)}Ni=0.
Q5. What is the similarity of Corollary 1 to Proposition 1?
The similarity of Corollary 1 to Proposition 1 stems from the estimation and control duality due to the differential nature of contraction analysis as is evident from (13) and (20).
Q6. What is the goal of the motion planning problem?
The goal of the motion planning problem is to find an optimal trajectory that avoids these obstacles and minimize ∫ 50 0 ‖u(t)‖2dt subject to input constraints 0 ≤ ui(t) ≤ 1,∀i,∀t and the dynamics constraints.
Q7. How many layers are underfit to the training samples?
The authors can see that the models with more than 2 layers overfit and those with less than 32 hidden units underfit to the training samples.
Q8. What is the simplest formulation for the NCM?
In this section, the authors propose its simpler formulation for nonlinear systems with bounded disturbances in order to be of practical use in engineering applications.
Q9. how can i find in the problem (8)?
although α is fixed in Theorem 2, it can be found by a line search as will be demonstrated in Section V.Remark 2: The problem (8) can be solved as a finitedimensional problem by using backward difference approximation, ˙̃W(x(ti), ti) (W̃(x(ti), ti)− W̃(x(ti−1), ti−1))/ ti, where t = ti − ti−1,∀i with t t2 > 0, and by discretizing it along a pre-computed system trajectory {x(ti)}Ni=0.
Q10. How do the authors get a contraction metric?
Multiplying (3) by μ−1 from both sides gives ωI W(x, t) ωI as the authors have ω,ω > 0 and (μ−1)2 = W. The authors get (7) by multiplying this inequality by ν = 1/ω.
Q11. What is the difference between the two definitenesses?
Instead of directly using sequential data of optimal contraction metrics {M(x(ti), ti)}Ni=0 for neural network training, the positive definiteness of M(x, t) is utilized to reduce the dimension of the target output {yi}Ni=0 defined in Section II.
Q12. What is the MSE of the NCM estimator?
As expected from the small MSE of Table I, Figure 4 implies that the NCM estimator is able to approximate the sampling-based CV-STEM estimator without losing its estimation performance.
Q13. What is the way to measure the optimal contraction metric?
1) Sampling of Optimal Contraction Metrics: Using Proposition 1, the authors sample the optimal contraction metric along 100 trajectories with uniformly distributed initial conditions (−10 ≤ xi ≤ 10, i = 1, 2, 3).