Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
167 tokens/sec
GPT-4o
7 tokens/sec
Gemini 2.5 Pro Pro
42 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
38 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Neural Contractive Dynamical Systems (2401.09352v1)

Published 17 Jan 2024 in cs.RO, cs.AI, and cs.LG

Abstract: Stability guarantees are crucial when ensuring a fully autonomous robot does not take undesirable or potentially harmful actions. Unfortunately, global stability guarantees are hard to provide in dynamical systems learned from data, especially when the learned dynamics are governed by neural networks. We propose a novel methodology to learn neural contractive dynamical systems, where our neural architecture ensures contraction, and hence, global stability. To efficiently scale the method to high-dimensional dynamical systems, we develop a variant of the variational autoencoder that learns dynamics in a low-dimensional latent representation space while retaining contractive stability after decoding. We further extend our approach to learning contractive systems on the Lie group of rotations to account for full-pose end-effector dynamic motions. The result is the first highly flexible learning architecture that provides contractive stability guarantees with capability to perform obstacle avoidance. Empirically, we demonstrate that our approach encodes the desired dynamics more accurately than the current state-of-the-art, which provides less strong stability guarantees.

Definition Search Book Streamline Icon: https://streamlinehq.com
References (44)
  1. Learning Riemannian manifolds for geodesic motion skills. In Robotics: Science and Systems (R:SS), 2021. URL https://roboticsconference.org/2021/program/papers/082/index.html.
  2. Learning from humans. In B. Siciliano and O. Khatib (eds.), Handbook of Robotics, chapter 74, pp.  1995–2014. Springer, Secaucus, NJ, USA, 2016. URL https://doi.org/10.1007/978-3-319-32552-1_74. 2nd Edition.
  3. Learning stable dynamical systems using contraction theory. In IEEE International Conference on Ubiquitous Robots and Ambient Intelligence (URAI), pp.  124–129, 2017. URL https://ieeexplore.ieee.org/document/7992901.
  4. Flows for simultaneous manifold learning and density estimation. In Neural Information Processing Systems (NeurIPS, pp.  442–453, 2020. URL https://proceedings.neurips.cc/paper_files/paper/2020/file/051928341be67dcba03f0e04104d9047-Paper.pdf.
  5. Active learning based on data uncertainty and model sensitivity. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp.  1547–1554, 2018a. URL http://doi.org/10.1109/IROS.2018.8593552.
  6. Ricky T. Q. Chen. torchdiffeq, 2018. URL https://github.com/rtqichen/torchdiffeq.
  7. Neural ordinary differential equations. In Neural Information Processing Systems (NeurIPS), 2018b. URL https://proceedings.neurips.cc/paper/2018/file/69386f6bb1dfed68692a24c8686939b9-Paper.pdf.
  8. Not your grandmother’s toolbox – the robotics toolbox reinvented for Python. In IEEE International Conference on Robotics and Automation (ICRA), pp.  11357–11363, 2021. URL https://10.1109/ICRA48506.2021.9561366.
  9. Safe control with learned certificates: A survey of neural Lyapunov, barrier, and contraction methods for robotics and control. IEEE Transactions on Robotics (T-RO), 39(3):1749–1767, 2023. URL https://doi.org/10.1109/TRO.2022.3232542.
  10. Reparameterizing distributions on Lie groups. In International Conference on Artificial Intelligence and Statistics (AISTATS), pp.  3244–3253, 2019. URL https://proceedings.mlr.press/v89/falorsi19a.html.
  11. Avoidance of convex and concave obstacles with convergence ensured through contraction. IEEE Robotics and Automation Letters (RA-L), 4(2):1462–1469, 2019. doi: http://doi.org/10.1109/LRA.2019.2893676.
  12. Avoiding dense and dynamic obstacles in enclosed spaces: Application to moving in crowds. IEEE Transactions on Robotics, 38(5):3113–3132, 2022. URL https://ieeexplore.ieee.org/document/9765824.
  13. Reaching through latent space: From joint statistics to path planning in manipulation. IEEE Robotics and Automation Letters (RA-L), 7(2):5334–5341, 2022. URL http://doi.org/10.1109/LRA.2022.3152697.
  14. A tutorial on incremental stability analysis using contraction theory. Modeling, Identification and Control, 2010. ISSN 0332-7353. URL https://doi.org/10.4173/mic.2010.3.2.
  15. Learning stable nonlinear dynamical systems with Gaussian mixture models. IEEE Transactions on Robotics, 27(5):943–957, 2011. doi: 10.1109/TRO.2011.2159412.
  16. Learning control Lyapunov function to ensure stability of dynamical system-based robot reaching motions. Robotics and Autonomous Systems (RAS), 62(6):752–765, 2014. doi: https://doi.org/10.1016/j.robot.2014.03.001.
  17. Auto-encoding variational Bayes. In International Conference on Learning Representations (ICLR), 2014. URL https://openreview.net/forum?id=33X9fd2-9FyZd.
  18. The kit bimanual manipulation dataset. In IEEE/RAS International Conference on Humanoid Robots (Humanoids), pp.  499–506, 2021. URL https://motion-database.humanoids.kit.edu/.
  19. Open-source benchmarking for learned reaching motion generation in robotics. Paladyn, Journal of Behavioral Robotics, 6(1), 2015. URL http://doi.org/10.1515/pjbr-2015-0002.
  20. On contraction analysis for non-linear systems. Automatica, 34(6):683–696, 1998. ISSN 0005-1098. URL https://doi.org/10.1016/S0005-1098(98)00019-3.
  21. JacNet: Learning functions with structured Jacobians. INNF Workshop at the International Conference on Machine Learning (ICML), 2019. URL https://invertibleworkshop.github.io/INNF_2019/accepted_papers/pdfs/INNF_2019_paper_10.pdf.
  22. Control contraction metrics: Convex and intrinsic criteria for nonlinear feedback design. IEEE Transactions on Automatic Control, 62(6):3046–3053, 2017. URL https://doi.org/10.1109/TAC.2017.2668380.
  23. Learning robot motions with stable dynamical systems under diffeomorphic transformations. Robotics and Autonomous Systems (RAS), 70:1–15, 2015. URL https://doi.org/10.1016/j.robot.2015.04.006.
  24. On second order behaviour in augmented neural ODEs: A short summary. In Workshop on the Symbiosis of Deep Learning and Differential Equations at NeurIPS, 2021. URL https://openreview.net/forum?id=XpmaGtI04ki.
  25. Pytorch: An imperative style, high-performance deep learning library, 2019. URL https://proceedings.neurips.cc/paper/2019/file/bdbca288fee7f92f2bfa9f7012727740-Paper.pdf.
  26. Torchdyn: Implicit models and neural numerical methods in pytorch. URL https://github.com/DiffEqML/torchdyn/.
  27. Euclideanizing flows: Diffeomorphic reduction for learning stable dynamical systems. In Conference on Learning for Dynamics and Control (L4DC), pp.  630–639, 2020. URL https://proceedings.mlr.press/v120/rana20a.html.
  28. Learning position and orientation dynamics from demonstrations via contraction analysis. Autonomous Robots, 43(4):897–912, 2019. ISSN 0929-5593. URL https://doi.org/10.1007/s10514-018-9758-x.
  29. Learning partially contracting dynamical systems from demonstrations. In Conference on Robot Learning (CoRL), pp.  369–378, 2017. URL https://proceedings.mlr.press/v78/ravichandar17a.html.
  30. Stochastic backpropagation and approximate inference in deep generative models. In International Conference on Machine Learning (ICML), pp.  1278–1286, 2014.
  31. Computational approaches to motor learning by imitation. Phil. Trans. R. Soc. Lond. B, 358:537–547, 2003. URL http://doi.org/10.1098/rstb.2002.1258.
  32. Malcolm David Shuster. A survey of attitude representation. Journal of The Astronautical Sciences, 41:439–517, 1993.
  33. Learning contracting vector fields for stable imitation learning. arXiv, 1804.04878, 2018. URL https://arxiv.org/abs/1804.04878.
  34. A micro lie theory for state estimation in robotics. CoRR, abs/1812.01537, 2018. URL http://arxiv.org/abs/1812.01537.
  35. Learning certified control using contraction metric. In Conference on Robot Learning (CoRL), 2020. URL http://arxiv.org/abs/2011.12569.
  36. A family of nonparametric density estimation algorithms. Communications on Pure and Applied Mathematics, 66(2):145–164, 2013.
  37. Neural contraction metrics for robust estimation and control: A convex optimization approach. IEEE Control Systems Letters, 5:211–216, 2021. ISSN 24751456. URL https://doi.org/10.1109/LCSYS.2020.3001646.
  38. A theoretical overview of neural contraction metrics for learning-based control with guaranteed stability. In IEEE Conference on Decision and Control (CDC), pp.  2949–2954, 2021a. URL https://ieeexplore.ieee.org/document/9682859.
  39. Contraction theory for nonlinear stability analysis and learning-based control: A tutorial overview. Annual Reviews in Control, 52:135–169, 2021b. ISSN 1367-5788. URL https://doi.org/10.1016/j.arcontrol.2021.10.001.
  40. Imitationflow: Learning deep stable stochastic dynamic systems by normalizing flows. In International Conference on Intelligent Robots and Systems (IROS), pp.  5231–5237. IEEE, 2020. URL http://ras.papercept.net/images/temp/IROS/files/1340.pdf.
  41. Learning stable vector fields on lie groups. IEEE Robotics and Automation Letters, 7(4):12569–12576, 2022. doi: 10.1109/LRA.2022.3219019. URL https://ieeexplore.ieee.org/document/9935105.
  42. How neural networks extrapolate: From feedforward to graph neural networks. In International Conference on Learning Representations (ICLR), 2021. URL https://openreview.net/forum?id=UH-cmocLJC.
  43. Learning Riemannian stable dynamical systems via diffeomorphisms. In Conference on Robot Learning (CoRL), 2022a. URL https://openreview.net/pdf?id=o8dLx8OVcNk.
  44. Learning accurate and stable point-to-point motions: A dynamic system approach. IEEE Robotics and Automation Letters (RA-L), 7(2):1510–1517, 2022b. URL https://doi.org/10.1109/LRA.2022.3140677.
Citations (2)
View on Semantic Scholar

Summary

  • The paper introduces a novel neural network architecture based on contraction theory that guarantees global stability across its entire parameter space.
  • It leverages a VAE-inspired approach to efficiently scale to high-dimensional systems, preserving contractive stability even for complex motions including rotations.
  • Empirical results demonstrate that NCDS outperforms existing models by accurately capturing dynamics and integrating obstacle avoidance within motion planning.

Introduction to Neural Contractive Dynamical Systems

Stability is a cornerstone in robotics, particularly when programming a robot to act autonomously. Without stability, robots might exhibit unreliable or even dangerous behaviors. Traditionally, programming stable robot behavior has involved using hand-crafted dynamics, an approach that’s both time-intensive and inflexible. However, the rise of machine learning has opened the door to learning robotic dynamics directly from data, though this method comes with its own set of challenges – chiefly, ensuring stability. In a novel approach, a new framework called Neural Contractive Dynamical Systems (NCDS) tackles this problem head-on.

Ensuring Stability Through Contraction

Contraction theory is the underpinning force behind ensuring stability in NCDS. It enables a robot to digest a set of demonstration behaviors and generalize them in a way that maintains consistency, even when faced with perturbations. Unlike traditional methods that can guarantee either asymptotic or contraction stability (the latter being stronger of the two), existing learning-based models struggle with offering consistent stability due to their neural network foundations. The proposed NCDS architecture, by design, ensures contraction stability across the entire parameter space.

NCDS Functionality at a Glance

The driving force of NCDS is a neural network architecture that inherently respects contraction theory's mandate for global stability. This is accomplished by constructing the network's Jacobian, or its transformation blueprint, to be contractive regardless of the parameter values. With this property, NCDS offers a versatile tool that can be used within existing frameworks to provide stability guarantees without the need for imposing external constraints during training – a notable advantage over previous methods.

Scaling to High-Dimensional Systems

One might question the applicability of such a system to complex, high-dimensional problems. NCDS addresses this by incorporating a variant of the Variational Autoencoder (VAE) designed to comprehend dynamics in a compressed, low-dimensional latent space. This method not only retains contractive stability upon decoding but is also extended to encompass full-pose movements including rotations, thanks to its operation on the special orthogonal group 𝒮𝒪(3).

Empirical Results and Applications

Empirical results display NCDS's superiority in capturing the desired dynamics with more accuracy compared to existing models. It also scales efficiently to high-dimensional systems and even demonstrates the ability to integrate obstacle avoidance into motion planning. This is showcased through experiments on synthetic and real datasets, including a 7-DoF robotic manipulator tasked with a drawing activity. The avoidance is modeled using dynamic modulation, ensuring that the robot evades obstacles while preserving the stability guarantees intrinsic to NCDS.

Concluding Remarks

NCDS marks a significant step toward reliable robot learning that doesn't compromise on stability. By harnessing contraction theory within its neural network design, NCDS offers a method to program autonomous robots that are stable, adaptable, and capable of learning from high-dimensional data. While certain limitations exist, such as the reliance on adaptive step sizing during numerical integration, the benefits of stability guarantees are an attractive trade-off for the robotic systems of the future.

X Twitter Logo Streamline Icon: https://streamlinehq.com

Tweets