pith. sign in

arxiv: 2311.03600 · v3 · submitted 2023-11-06 · 💻 cs.RO

Scalable and Efficient Continual Learning from Demonstration via a Hypernetwork-generated Stable Dynamics Model

Sayantan Auddy , Jakob Hollenstein , Matteo Saveriano , Antonio Rodr\'iguez-S\'anchez , Justus Piater This is my paper

Pith reviewed 2026-05-24 05:41 UTC · model grok-4.3

classification 💻 cs.RO
keywords continual learninglearning from demonstrationstable dynamicshypernetworksneural ODELyapunov stabilityrobot motion skills
0
0 comments X

The pith

A hypernetwork generates parameters for a stable neural ODE that lets robots learn sequences of motion skills from demonstration without forgetting earlier ones and with linear total training time.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that a hypernetwork can output the weights of both a dynamics network and a Lyapunov function to form a clock-augmented stable neural ODE solver. This model is trained sequentially on new demonstration tasks while a stochastic regularization term applied to a single uniformly sampled task embedding keeps prior skills intact. The resulting system reports better trajectory accuracy and stability scores than prior stable LfD methods across datasets with up to 26 tasks and 32-dimensional trajectories. Stability of the generated dynamics is presented as a direct contributor to the observed continual-learning performance. The approach is evaluated on both synthetic high-dimensional LASA variants and real robot position-plus-orientation tasks.

Core claim

The central claim is that a hypernetwork can produce the full parameter set of a clock-augmented sNODE consisting of a trajectory-learning neural ODE and a trajectory-stabilizing Lyapunov function, and that adding stochastic regularization over a single uniformly sampled task embedding is sufficient to learn N tasks sequentially, reducing cumulative training cost from quadratic to linear in N while preserving stability guarantees and without degrading real-world performance.

What carries the argument

Hypernetwork-generated clock-augmented sNODE: the hypernetwork maps a task embedding to the weights of both the neural ODE dynamics model and its associated Lyapunov function, with an explicit clock input that enables stable forward integration of demonstrated trajectories.

If this is right

  • Robots can acquire and execute long sequences of motion skills from demonstration without storing or retraining on past data.
  • Total training cost for N skills scales linearly rather than quadratically.
  • Stability of the learned dynamics measurably improves continual-learning metrics, especially inside compact chunked hypernetworks.
  • The same architecture handles trajectories from 2 to 32 dimensions and real position-plus-orientation robot tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The single-embedding regularization could support smooth interpolation between learned skills if the embedding space is treated as continuous.
  • The stability-continual-learning link may transfer to other dynamical-system domains that require guaranteed convergence.
  • Extending the clock input to variable-speed or event-triggered clocks could enlarge the class of admissible trajectories without retraining the hypernetwork.

Load-bearing premise

A single uniformly sampled task embedding plus stochastic regularization is sufficient to block interference across tasks while preserving the stability guarantees of every previously generated Lyapunov function.

What would settle it

A sequence of 20 or more tasks in which either total training time grows quadratically, or stability metrics degrade, or trajectory reproduction error rises above the reported baselines would falsify the O(N) claim and the non-interference guarantee.

Figures

Figures reproduced from arXiv: 2311.03600 by Antonio Rodr\'iguez-S\'anchez, Jakob Hollenstein, Justus Piater, Matteo Saveriano, Sayantan Auddy.

Figure 1
Figure 1. Figure 1: Overview of key results and our proposed approach. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Time dependent stable NODE (sNODE) architecture: time input is added to the sNODE resulting in more accurate predictions (changes are shown in purple.). This change is done to enable the combination of ˆfθ(xˆt) with the gradient of the Lyapunov function (in Eq. (4)), which is now defined as ∇V (xˆt) =  ∂V ∂x0t , ∂V ∂x1t , · · · ∂V ∂xn−1t , ∂V ∂t T (11) Since the Lyapunov function Vγ produces a scalar val… view at source ↗
Figure 3
Figure 3. Figure 3: Trajectories of position and orientation are learned simultaneously by [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (a) Architecture of the HN→sNODE model. Parameters θ and γ of the nominal dynamics model ˆfθ and the Lyapunov function Vγ, respectively, of the sNODE are generated by the final layer of the Hypernetwork. (b) Architecture of the CHN→sNODE model. Parameters θ and γ of the sNODE are generated in chunks by the Chunked Hypernetwork, allowing for a smaller hypernetwork size. For (a) and (b), the architecture of … view at source ↗
Figure 5
Figure 5. Figure 5: DTW errors (lower is better) of all predictions while learning the [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: sNODE improves continual learning performance compared to NODE : DTW errors (lower is better) of trajectories predicted by SG, REP, HN, and CHN while learning the tasks of the LASA dataset are shown. The x-axis shows the current task. After learning each new task, the current and all previous tasks are evaluated. The DTW errors of these predictions are shown on the y-axis. Points show the medians and the s… view at source ↗
Figure 7
Figure 7. Figure 7: Parameter counts after learning all 26 tasks of the LASA 2D dataset. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Continual learning metrics (0:worst-1:best) for SG, REP, HN, and [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative examples of predictions made for the LASA 2D dataset. CHN [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative examples of predictions produced by CHN [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Comparison of the DTW errors (lower is better) of all predictions [PITH_FULL_IMAGE:figures/full_fig_p011_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: DTW errors of trajectories predicted by SG, REP, HN, and CHN while learning the tasks of the LASA 32D dataset (lower is better). The x-axis [PITH_FULL_IMAGE:figures/full_fig_p012_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Comparison of the DTW errors (lower is better) for different [PITH_FULL_IMAGE:figures/full_fig_p012_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: 2D boxplots showing the position errors (DTW) in the y-axis, and orientation errors (quaternion error) in the x-axis of all predictions while learning [PITH_FULL_IMAGE:figures/full_fig_p013_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Position (top) and orientation (bottom) errors of SG, REP, HN, and CHN for the RoboTasks9 dataset (lower is better). The x-axis shows the [PITH_FULL_IMAGE:figures/full_fig_p013_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Continual learning metrics (0:worst-1:best) for SG, REP, HN, and [PITH_FULL_IMAGE:figures/full_fig_p013_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Comparison of continual learning metrics (0:worst-1:best) for HN, [PITH_FULL_IMAGE:figures/full_fig_p014_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Parameter counts after learning all 9 tasks of the RoboTasks9 dataset. [PITH_FULL_IMAGE:figures/full_fig_p014_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Stochastic regularization in CHN→sNODE for the RoboTasks9 dataset: (top) position errors, (middle) orientation errors, and (bottom) training time for each task. The upper baseline SG (using sNODE) provides the reference for good performance, but has many more parameters than the CHN models. CHN-1 (ours), CHN-3, and CHN-5 use 1, 3 and 5 randomly selected task embeddings for regularization respectively. CHN… view at source ↗
Figure 21
Figure 21. Figure 21: DTW errors (y-axis) for stochastic regularization in CHN [PITH_FULL_IMAGE:figures/full_fig_p015_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Qualitative examples for the RoboTasks9 dataset. Models of CHN [PITH_FULL_IMAGE:figures/full_fig_p016_22.png] view at source ↗
read the original abstract

Robots capable of learning from demonstration (LfD) must exhibit stability while executing learned motion skills. To be effective in the real world, they should also remember multiple skills over time -- a capability lacking in current stable-LfD methods. We propose an approach to stable, continual LfD, and highlight the role of stability in improving continual learning. Our proposed hypernetwork generates the parameters of two neural networks: a trajectory learning dynamics model, and a trajectory-stabilizing Lyapunov function. These generated networks form a clock-augmented stable neural ODE solver (sNODE), a stable dynamics model that offers a superior stability-accuracy trade-off compared to the state-of-the-art. We further propose stochastic hypernetwork regularization with a single, uniformly-sampled task embedding, reducing the cumulative training time for $N$ tasks from O($N^2$) to O($N$) without degrading performance on real-world tasks. We introduce high-dimensional variants of the popular LASA dataset to assess scalability and extend a dataset of robotic LfD tasks to assess real-world performance. We empirically evaluate our approach on multiple LfD datasets of varying complexity, including sequences of 7--26 tasks, trajectories of 2--32 dimensions, and real-world tasks involving position and orientation. Our thorough evaluation on multiple LfD datasets demonstrates that our approach sequentially learns and retains multiple motion skills without retraining on past demonstrations, and outperforms other relevant baselines in terms of trajectory errors, continual learning scores, and stability metrics. Notably, we show that stability greatly enhances continual learning performance, particularly in size-efficient chunked hypernetworks. Our code is available at https://github.com/sayantanauddy/clfd-snode.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper claims to introduce a hypernetwork that generates parameters for both a dynamics model and a Lyapunov function, forming a clock-augmented stable neural ODE (sNODE) for continual learning from demonstration. It proposes stochastic hypernetwork regularization with a single uniformly-sampled task embedding to reduce cumulative training time for N tasks from O(N²) to O(N). Evaluations on high-dimensional LASA variants and extended robotic LfD datasets (sequences of 7-26 tasks, 2-32 dimensions) show outperformance over baselines in trajectory errors, continual learning scores, and stability metrics, with stability enhancing continual learning performance, particularly for chunked hypernetworks. Code is released.

Significance. If the Lyapunov stability certificates remain valid for prior tasks after hypernetwork updates, the work would represent a notable advance in scalable stable LfD by addressing catastrophic forgetting while maintaining stability guarantees. The O(N) scaling via regularization, empirical superiority on real-world position/orientation tasks, and released code for reproducibility strengthen its potential impact in robotics and continual learning.

major comments (2)
  1. The O(N) training claim and retention of stability without retraining rest on the assumption that stochastic regularization with one randomly drawn task embedding suffices to preserve the Lyapunov decrease condition for all prior tasks. The sequential training protocol (7-26 tasks) reports only final aggregate metrics and does not isolate or verify whether the generated V still satisfies stability for earlier embeddings after later updates, which is load-bearing for the central continual-learning claims.
  2. The abstract and method description reference stability proofs and a Lyapunov construction generated by the hypernetwork, but it is unclear whether these derivations explicitly account for weight changes induced by updates on new task embeddings; any gaps here would invalidate the stability guarantees for previously learned skills.
minor comments (1)
  1. The abstract could more precisely state the exact task counts, dimensions, and dataset variants used in the LASA and robotic evaluations to improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed and constructive comments, which highlight important aspects of the stability guarantees in our continual learning approach. We respond to each major comment below and indicate planned revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: The O(N) training claim and retention of stability without retraining rest on the assumption that stochastic regularization with one randomly drawn task embedding suffices to preserve the Lyapunov decrease condition for all prior tasks. The sequential training protocol (7-26 tasks) reports only final aggregate metrics and does not isolate or verify whether the generated V still satisfies stability for earlier embeddings after later updates, which is load-bearing for the central continual-learning claims.

    Authors: We agree that the current results present only aggregate metrics and do not include explicit per-task verification of the Lyapunov decrease condition after subsequent updates. The stochastic regularization samples a single task embedding uniformly at each step to encourage preservation of stability properties across the task distribution without incurring O(N^2) cost. To directly address this point, the revised manuscript will include additional analysis that evaluates the Lyapunov condition on earlier task embeddings after training proceeds to later tasks, thereby isolating the effect of the regularization. revision: yes

  2. Referee: The abstract and method description reference stability proofs and a Lyapunov construction generated by the hypernetwork, but it is unclear whether these derivations explicitly account for weight changes induced by updates on new task embeddings; any gaps here would invalidate the stability guarantees for previously learned skills.

    Authors: The stability certificates are established by construction for each pair of dynamics and Lyapunov networks generated by the hypernetwork for a fixed task embedding; the relevant derivations show that the generated sNODE satisfies the Lyapunov decrease condition at generation time. Updates to the hypernetwork are performed under the stochastic regularization objective, which is intended to keep previously seen embeddings within the region where the generated Lyapunov functions remain valid. We will revise the method section to explicitly distinguish the per-embedding stability guarantee from the effect of hypernetwork updates and to clarify how the regularization objective supports retention of those guarantees. revision: partial

Circularity Check

0 steps flagged

No significant circularity; claims rest on empirical evaluation and explicit regularization design

full rationale

The paper presents an sNODE construction via hypernetwork-generated dynamics and Lyapunov function, plus a stochastic regularization scheme using one uniformly sampled task embedding. These are architectural choices whose O(N) scaling and stability properties are asserted as direct consequences of the design and then validated empirically on LASA variants and real-robot tasks, with code released. No derivation step reduces a claimed prediction or uniqueness result to a fitted parameter or self-citation by construction; the central performance claims are benchmark comparisons rather than tautological re-derivations of inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach relies on the existence of a Lyapunov function that can be generated by the hypernetwork to guarantee stability for each task; no explicit free parameters are named in the abstract, but the task embeddings and hypernetwork architecture choices function as design parameters.

axioms (1)
  • domain assumption A Lyapunov function generated alongside the dynamics model guarantees asymptotic stability of the learned trajectories.
    Invoked to justify the stability-accuracy trade-off claim.

pith-pipeline@v0.9.0 · 5862 in / 1399 out tokens · 27953 ms · 2026-05-24T05:41:21.580897+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Continual Domain Randomization

    cs.RO 2024-03 unverdicted novelty 6.0

    Continual Domain Randomization trains RL policies sequentially on randomization parameter subsets with continual learning to achieve robust sim-to-real transfer in robotic reaching and grasping.

Reference graph

Works this paper leans on

59 extracted references · 59 canonical work pages · cited by 1 Pith paper · 5 internal anchors

  1. [1]

    Recent advances in robot learning from demonstration,

    H. Ravichandar, A. S. Polydoros, S. Chernova, and A. Billard, “Recent advances in robot learning from demonstration,” Annual review of control, robotics, and autonomous systems , vol. 3, pp. 297–330, 2020

    work page 2020

  2. [2]

    Imitationflow: Learning deep stable stochastic dynamic systems by normalizing flows,

    J. Urain, M. Ginesi, D. Tateo, and J. Peters, “Imitationflow: Learning deep stable stochastic dynamic systems by normalizing flows,” in 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2020, pp. 5231–5237

    work page 2020

  3. [3]

    Dynamical system modulation for robot learning via kinesthetic demonstrations,

    M. Hersch, F. Guenter, S. Calinon, and A. Billard, “Dynamical system modulation for robot learning via kinesthetic demonstrations,” IEEE Transactions on Robotics , vol. 24, no. 6, pp. 1463–1467, 2008

    work page 2008

  4. [4]

    An energy-based approach to ensure the stability of learned dynamical systems,

    M. Saveriano, “An energy-based approach to ensure the stability of learned dynamical systems,” in IEEE International Conference on Robotics and Automation (ICRA) , 2020, pp. 4407–4413

    work page 2020

  5. [5]

    Learning control lyapunov function to ensure stability of dynamical system-based robot reaching motions,

    S. M. Khansari-Zadeh and A. Billard, “Learning control lyapunov function to ensure stability of dynamical system-based robot reaching motions,” Robotics and Autonomous Systems, vol. 62, no. 6, pp. 752–765, 2014

    work page 2014

  6. [6]

    Learning stable deep dynamics models,

    J. Z. Kolter and G. Manek, “Learning stable deep dynamics models,” Advances in Neural Information Processing Systems , vol. 32, pp. 11 128– 11 136, 2019

    work page 2019

  7. [7]

    Continual learning from demonstration of robotics skills,

    S. Auddy, J. Hollenstein, M. Saveriano, A. Rodríguez-Sánchez, and J. Piater, “Continual learning from demonstration of robotics skills,” Robotics and Autonomous Systems , vol. 165, p. 104427, 2023. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ S0921889023000660

    work page 2023

  8. [8]

    Neural ordinary differential equations,

    R. T. Chen, Y . Rubanova, J. Bettencourt, and D. Duvenaud, “Neural ordinary differential equations,” in Proceedings of the 32nd International Conference on Neural Information Processing Systems , 2018, pp. 6572– 6583

    work page 2018

  9. [9]

    Hypernetworks,

    D. Ha, A. M. Dai, and Q. V . Le, “Hypernetworks,” in International Conference on Learning Representations , 2017. [Online]. Available: https://openreview.net/forum?id=rkpACe1lx

    work page 2017

  10. [10]

    Continual learning with hypernetworks,

    J. von Oswald, C. Henning, J. Sacramento, and B. F. Grewe, “Continual learning with hypernetworks,” in International Conference on Learning Representations (ICLR), 2019

    work page 2019

  11. [11]

    Learning stable nonlinear dynamical systems with Gaussian mixture models,

    S. M. Khansari-Zadeh and A. Billard, “Learning stable nonlinear dynamical systems with Gaussian mixture models,” IEEE Transactions on Robotics, vol. 27, no. 5, pp. 943–957, 2011

    work page 2011

  12. [12]

    Continual lifelong learning with neural networks: A review,

    G. I. Parisi, R. Kemker, J. L. Part, C. Kanan, and S. Wermter, “Continual lifelong learning with neural networks: A review,” Neural Networks, vol. 113, pp. 54–71, 2019

    work page 2019

  13. [13]

    Learning from demonstration (programming by demonstra- tion),

    S. Calinon, “Learning from demonstration (programming by demonstra- tion),” Encyclopedia of robotics , pp. 1–8, 2018

    work page 2018

  14. [14]

    Learning from humans,

    A. Billard, S. Calinon, and R. Dillmann, “Learning from humans,” Springer Handbook of Robotics, 2nd Ed. , 2016. 18

    work page 2016

  15. [15]

    A survey of robot learning from demonstration,

    B. D. Argall, S. Chernova, M. Veloso, and B. Browning, “A survey of robot learning from demonstration,” Robotics and autonomous systems , vol. 57, no. 5, pp. 469–483, 2009

    work page 2009

  16. [16]

    Robot learning from demonstration: A review of recent advances,

    H. Ravichandar, A. Polydoros, S. Chernova, and A. Billard, “Robot learning from demonstration: A review of recent advances,” Annual Review of Control, Robotics, and Autonomous Systems , 2019

    work page 2019

  17. [17]

    Trajectory-based skill learning using generalized cylinders,

    S. R. Ahmadzadeh and S. Chernova, “Trajectory-based skill learning using generalized cylinders,” Frontiers in Robotics and AI , vol. 5, p. 132, 2018

    work page 2018

  18. [18]

    CRIL: Continual robot imitation learning via generative and prediction model,

    C. Gao, H. Gao, S. Guo, T. Zhang, and F. Chen, “CRIL: Continual robot imitation learning via generative and prediction model,” in 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021, pp. 6747–5754

    work page 2021

  19. [19]

    Towards one shot learning by imitation for humanoid robots,

    Y . Wu and Y . Demiris, “Towards one shot learning by imitation for humanoid robots,” in 2010 IEEE international conference on robotics and automation. IEEE, 2010, pp. 2889–2894

    work page 2010

  20. [20]

    Teacher feedback to scaffold and refine demonstrated motion primitives on a mobile robot,

    B. D. Argall, B. Browning, and M. M. Veloso, “Teacher feedback to scaffold and refine demonstrated motion primitives on a mobile robot,” Robotics and Autonomous Systems , vol. 59, no. 3-4, pp. 243–255, 2011

    work page 2011

  21. [21]

    Inverse kkt: Learning cost functions of manipulation tasks from demonstrations,

    P. Englert, N. A. Vien, and M. Toussaint, “Inverse kkt: Learning cost functions of manipulation tasks from demonstrations,” The International Journal of Robotics Research , vol. 36, no. 13-14, pp. 1474–1488, 2017

    work page 2017

  22. [22]

    Compliant skills acquisition and multi-optima policy search with em-based reinforcement learning,

    S. Calinon, P. Kormushev, and D. G. Caldwell, “Compliant skills acquisition and multi-optima policy search with em-based reinforcement learning,” Robotics and Autonomous Systems, vol. 61, no. 4, pp. 369–379, 2013

    work page 2013

  23. [23]

    Model-based inverse reinforcement learning from visual demonstrations,

    N. Das, S. Bechtle, T. Davchev, D. Jayaraman, A. Rai, and F. Meier, “Model-based inverse reinforcement learning from visual demonstrations,” in Conference on Robot Learning . PMLR, 2021, pp. 1930–1942

    work page 2021

  24. [24]

    Learning stable robotic skills on riemannian manifolds,

    M. Saveriano, F. J. Abu-Dakka, and V . Kyrki, “Learning stable robotic skills on riemannian manifolds,” Robotics and Autonomous Systems , vol. 169, p. 104510, 2023

    work page 2023

  25. [25]

    Se (3)-diffusionfields: Learning smooth cost functions for joint grasp and motion optimization through diffusion,

    J. Urain, N. Funk, J. Peters, and G. Chalvatzaki, “Se (3)-diffusionfields: Learning smooth cost functions for joint grasp and motion optimization through diffusion,” in 2023 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2023, pp. 5923–5930

    work page 2023

  26. [26]

    A physically-consistent bayesian non-parametric mixture model for dynamical system learning,

    N. B. Figueroa Fernandez and A. Billard, “A physically-consistent bayesian non-parametric mixture model for dynamical system learning,” Proceedings of Machine Learning Research , 2018

    work page 2018

  27. [27]

    Movement imitation with nonlinear dynamical systems in humanoid robots,

    A. J. Ijspeert, J. Nakanishi, and S. Schaal, “Movement imitation with nonlinear dynamical systems in humanoid robots,” in International Conference on Robotics and Automation (ICRA) , 2002, pp. 1398–1403

    work page 2002

  28. [28]

    Progressive Neural Networks

    A. A. Rusu, N. C. Rabinowitz, G. Desjardins, H. Soyer, J. Kirkpatrick, K. Kavukcuoglu, R. Pascanu, and R. Hadsell, “Progressive neural networks,” arXiv preprint arXiv:1606.04671 , 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  29. [29]

    icarl: Incremental classifier and representation learning,

    S.-A. Rebuffi, A. Kolesnikov, G. Sperl, and C. H. Lampert, “icarl: Incremental classifier and representation learning,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2017, pp. 2001–2010

    work page 2017

  30. [30]

    Continual learning with deep generative replay,

    H. Shin, J. K. Lee, J. Kim, and J. Kim, “Continual learning with deep generative replay,” in Proceedings of the 31st International Conference on Neural Information Processing Systems , 2017, pp. 2994–3003

    work page 2017

  31. [31]

    Overcoming catastrophic forgetting in neural networks,

    J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska et al., “Overcoming catastrophic forgetting in neural networks,” Proceedings of the national academy of sciences , vol. 114, no. 13, pp. 3521–3526, 2017

    work page 2017

  32. [32]

    Continual learning through synaptic intelligence,

    F. Zenke, B. Poole, and S. Ganguli, “Continual learning through synaptic intelligence,” in International Conference on Machine Learning . PMLR, 2017, pp. 3987–3995

    work page 2017

  33. [33]

    Memory aware synapses: Learning what (not) to forget,

    R. Aljundi, F. Babiloni, M. Elhoseiny, M. Rohrbach, and T. Tuytelaars, “Memory aware synapses: Learning what (not) to forget,” in Proceedings of the European Conference on Computer Vision (ECCV) , 2018, pp. 139–154

    work page 2018

  34. [34]

    A continual learning survey: Defying forgetting in classification tasks,

    M. Delange, R. Aljundi, M. Masana, S. Parisot, X. Jia, A. Leonardis, G. Slabaugh, and T. Tuytelaars, “A continual learning survey: Defying forgetting in classification tasks,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2021

    work page 2021

  35. [35]

    Lifelong robot learning,

    S. Thrun and T. M. Mitchell, “Lifelong robot learning,” Robotics and Autonomous Systems, vol. 15, no. 1, pp. 25–46, Jul. 1995

    work page 1995

  36. [36]

    Continual Learning for Affec- tive Robotics: Why, What and How?

    N. Churamani, S. Kalkan, and H. Gunes, “Continual Learning for Affec- tive Robotics: Why, What and How?” in 2020 29th IEEE International Conference on Robot and Human Interactive Communication (RO-MAN) . Naples, Italy: IEEE, Aug. 2020, pp. 425–431

    work page 2020

  37. [37]

    Continual Learning for Affective Robotics: A Proof of Concept for Wellbeing,

    N. Churamani, M. Axelsson, A. Çaldır, and H. Gunes, “Continual Learning for Affective Robotics: A Proof of Concept for Wellbeing,” in 2022 10th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIIW) . Nara, Japan: IEEE, Oct. 2022, pp. 1–8

    work page 2022

  38. [38]

    A Lifelong Learning Approach to Mobile Robot Navigation,

    B. Liu, X. Xiao, and P. Stone, “A Lifelong Learning Approach to Mobile Robot Navigation,” IEEE Robotics and Automation Letters , vol. 6, no. 2, pp. 1090–1096, Apr. 2021. [Online]. Available: https://ieeexplore.ieee.org/document/9345478/

    work page arXiv 2021

  39. [39]

    Gradient episodic memory for continual learning,

    D. Lopez-Paz and M. Ranzato, “Gradient episodic memory for continual learning,” Advances in neural information processing systems , vol. 30, 2017

    work page 2017

  40. [40]

    Development of a Framework for Continual Learning in Industrial Robotics,

    M. Trinh, J. Moon, L. Grundel, V . Hankemeier, S. Storms, and C. Brecher, “Development of a Framework for Continual Learning in Industrial Robotics,” in 2022 IEEE 27th International Conference on Emerging Technologies and Factory Automation (ETFA) . Stuttgart, Germany: IEEE, Sep. 2022, pp. 1–8. [Online]. Available: https://ieeexplore.ieee.org/document/9921432/

    work page arXiv 2022

  41. [41]

    Online Continual Learning for Control of Mobile Robots,

    A. Sarabakha, Z. Qiao, S. Ramasamy, and P. N. Suganthan, “Online Continual Learning for Control of Mobile Robots,” in 2023 International Joint Conference on Neural Networks (IJCNN) . Gold Coast, Australia: IEEE, Jun. 2023, pp. 1–10. [Online]. Available: https://ieeexplore.ieee.org/document/10191188/

    work page arXiv 2023

  42. [42]

    Continual learning with tiny episodic memories,

    A. Chaudhry, M. Rohrbach, M. Elhoseiny, T. Ajanthan, P. Dokania, P. Torr, and M. Ranzato, “Continual learning with tiny episodic memories,” in Workshop on Multi-Task and Lifelong Reinforcement Learning , 2019

    work page 2019

  43. [43]

    Learning without forgetting,

    Z. Li and D. Hoiem, “Learning without forgetting,” IEEE transactions on pattern analysis and machine intelligence, vol. 40, no. 12, pp. 2935–2947, 2017

    work page 2017

  44. [44]

    Efficient Lifelong Learning with A-GEM

    A. Chaudhry, M. Ranzato, M. Rohrbach, and M. Elhoseiny, “Efficient lifelong learning with a-gem,” arXiv preprint arXiv:1812.00420 , 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  45. [45]

    Riemannian walk for incremental learning: Understanding forgetting and intransi- gence,

    A. Chaudhry, P. K. Dokania, T. Ajanthan, and P. H. Torr, “Riemannian walk for incremental learning: Understanding forgetting and intransi- gence,” in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 532–547

    work page 2018

  46. [46]

    Continual model-based reinforcement learning with hypernetworks,

    Y . Huang, K. Xie, H. Bharadhwaj, and F. Shkurti, “Continual model-based reinforcement learning with hypernetworks,” in 2021 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2021, pp. 799–805

    work page 2021

  47. [47]

    Hypernetwork-ppo for continual reinforcement learning,

    P. Schöpf, S. Auddy, J. Hollenstein, and A. Rodriguez-Sanchez, “Hypernetwork-ppo for continual reinforcement learning,” in Deep Reinforcement Learning Workshop NeurIPS , 2022

    work page 2022

  48. [48]

    Proximal Policy Optimization Algorithms

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Prox- imal policy optimization algorithms,” arXiv preprint arXiv:1707.06347 , 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  49. [49]

    Input convex neural networks,

    B. Amos, L. Xu, and J. Z. Kolter, “Input convex neural networks,” in International Conference on Machine Learning . PMLR, 2017, pp. 146–155

    work page 2017

  50. [50]

    Orientation in cartesian space dynamic movement primitives,

    A. Ude, B. Nemec, T. Petri ´c, and J. Morimoto, “Orientation in cartesian space dynamic movement primitives,” in 2014 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2014, pp. 2997–3004

    work page 2014

  51. [51]

    Toward orientation learning and adaptation in cartesian space,

    Y . Huang, F. J. Abu-Dakka, J. Silvério, and D. G. Caldwell, “Toward orientation learning and adaptation in cartesian space,” IEEE Transactions on Robotics, vol. 37, no. 1, pp. 82–98, 2020

    work page 2020

  52. [52]

    Merging position and orientation motion primitives,

    M. Saveriano, F. Franzel, and D. Lee, “Merging position and orientation motion primitives,” in 2019 International Conference on Robotics and Automation (ICRA). IEEE, 2019, pp. 7041–7047

    work page 2019

  53. [53]

    Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting,

    X. Li, Y . Zhou, T. Wu, R. Socher, and C. Xiong, “Learn to grow: A continual structure learning framework for overcoming catastrophic forgetting,” in International Conference on Machine Learning . PMLR, 2019, pp. 3925–3934

    work page 2019

  54. [54]

    Lifelong Learning with Dynamically Expandable Networks

    J. Yoon, E. Yang, J. Lee, and S. J. Hwang, “Lifelong learning with dynamically expandable networks,” arXiv preprint arXiv:1708.01547 , 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  55. [55]

    Introduction to smooth manifolds

    J. M. Lee, “Introduction to smooth manifolds.” Springer, 2012

    work page 2012

  56. [56]

    Hypernetworks for continual semi-supervised learning,

    D. Brahma, V . K. Verma, and P. Rai, “Hypernetworks for continual semi-supervised learning,” arXiv preprint arXiv:2110.01856 , 2021

    work page arXiv 2021

  57. [57]

    Learning stable dynamical systems using contraction theory,

    C. Blocher, M. Saveriano, and D. Lee, “Learning stable dynamical systems using contraction theory,” in 2017 14th International Conference on Ubiquitous Robots and Ambient Intelligence (URAI) . IEEE, 2017, pp. 124–129

    work page 2017

  58. [58]

    Similarity measures for identifying material parameters from hysteresis loops using inverse analysis,

    C. F. Jekel, G. Venter, M. P. Venter, N. Stander, and R. T. Haftka, “Similarity measures for identifying material parameters from hysteresis loops using inverse analysis,” International Journal of Material Forming , vol. 12, no. 3, pp. 355–378, 2019

    work page 2019

  59. [59]

    Don't forget, there is more than forgetting: new metrics for Continual Learning

    N. Díaz-Rodríguez, V . Lomonaco, D. Filliat, and D. Maltoni, “Don’t forget, there is more than forgetting: new metrics for continual learning,” arXiv preprint arXiv:1810.13166 , 2018. 19 APPENDIX A. Stable NODE with Time input We present the benefit of introducing the additional time input to the sNODE model, as described in Sec. IV-A. For this, we train ...

    work page internal anchor Pith review Pith/arXiv arXiv 2018