Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Meta-learning with one context vector per task factor—updating only the vector for the factor that changed—yields OOD robustness and zero-shot context recombination, shown in sine regression and quadruped stair climbing.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

DMCM learns one context vector per task factor and updates only the relevant vector, improving out-of-distribution robustness in sine regression and quadruped locomotion via context sharing.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A useful CAVIA extension with a promising real-world quadruped result, but the sine OOD gain isn't yet pinned to disentanglement rather than the task curriculum; worth refereeing for the robot work alone. the 3 major comments →

arxiv 2509.01297 v1 pith:PQDX4TSI submitted 2025-09-01 cs.RO

Disentangled Multi-Context Meta-Learning: Unlocking robust and Generalized Task Learning

Seonsoo Kim , Jun-Gill Kang , Taehong Kim , Seongil Hong This is my paper
classification cs.RO
keywords meta-learningdisentangled representation learningcontext vectorsfactors of variationout-of-distribution generalizationquadruped locomotionsim-to-real transferzero-shot adaptation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper claims that a large share of meta-learning's OOD failures comes from forcing all task variation into one entangled context or parameter vector. DMCM instead keeps one small context vector per declared factor of variation, and it sequences training tasks so consecutive tasks differ in exactly one factor; the inner loop then updates only that factor's vector. The payoff the paper argues for is double: the model can tell which thing changed, so it stays stable when combinations of factors were missing from training, and contexts learned on different tasks can be spliced together to predict or act in unseen factor combinations without any adaptation. The headline evidence is a real Go1 quadruped climbing 17 cm stairs under low Kp and added payload, using a terrain context from simulation stair data and a robot-property context from 20 seconds of real flat-terrain walking—while the single-context baseline fails completely. In sine regression the same mechanism beats MAML, ANIL, and CAVIA under range-exclusion OOD tests and gives zero-shot predictions by recombining amplitude and phase contexts.

Core claim

DMCM is a meta-learning algorithm built on CAVIA's context-parameter idea but with K separate context vectors instead of one. Each vector is assigned to a declared factor of variation (amplitude vs phase; terrain vs robot-specific properties). Task sampling is constrained so that consecutive tasks share all factors except one, and the inner loop updates only the context vector belonging to that changed factor. After B warm-up tasks, the outer loop updates shared weights, and an optional recombination loop trains the network to work with context vectors that were never adapted together, which the paper reports is necessary for zero-shot recombination (loss drops from 3.66 to 0.069 with the lo

What carries the argument

The central object is the set of K disentangled context vectors {φ1,...,φK}, each initialized to zero and concatenated into the network (in the robot task, 20 parameters per vector at layer inputs). The load-bearing mechanism is the conditional task sampler p(Ti | Ti−1) in Algorithm 1: it guarantees that the only difference between consecutive tasks is the single labeled factor s, so the inner-loop gradient ∇φs is, by construction, attributed to that factor alone; the other vectors are frozen. The recombination loop then teaches the shared network to accept independently adapted vectors by computing a meta-gradient on test data drawn with stored context vectors from K−1 different past adapta

Load-bearing premise

The entire scheme rests on being given, for every training task, a correct hand-labeled grouping of task variation into separate factors, and on being able to sequence tasks so consecutive tasks differ in exactly one of those factors; if real variation mixes factors (low gains plus payload plus rough terrain) or the labels misattribute a change, the selective inner-loop updates cannot align context vectors with the true causes.

What would settle it

Train DMCM with the two factors deliberately confounded (e.g., in the sine task sample amplitude and phase from a joint distribution that makes low amplitude almost always co-occur with a particular phase, or in the robot task always pair high payload with low Kp), then evaluate on the anti-correlated OOD combinations. If the model's loss or policy success does not degrade relative to the factor-balanced training case, or if swapping one context vector while holding the other fixed fails to change predictions/behavior in the factor-consistent direction, then the selective-update mechanism is n

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Under the paper's results, robustness to missing factor combinations does not require seeing all combinations: with 40–80% of amplitude–phase ranges excluded from sine training, DMCM's loss stays lower and less variable than MAML/ANIL/CAVIA.
  • Context vectors become portable assets: a terrain context from one source and a robot context from another can be combined at deployment, so real stair climbing can be achieved with no real stair data for the robot-property context.
  • A policy trained with contexts that are randomly mixed across tasks sharing the same factor outperforms a policy given exact task-matched contexts under OOD conditions (simulation success 602 vs 151 in OOD robot properties), implying that context diversity during RL training is itself a regularizer.
  • Correct-factor context combinations consistently beat wrong-factor combinations in dynamics prediction, giving a diagnostic handle on sim-to-real discrepancies: mismatches between sim and real show up mainly in the robot-specific context, not the terrain context.
  • If the disentanglement is real, meta-learning's interpretability increases: body height and front-foot probing behavior shift in predictable directions when payload or terrain contexts are deliberately wrong.

Where Pith is reading between the lines

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

  • My inference: the actual disentanglement engine is the one-factor-at-a-time task sequencing, not the network architecture; if that is right, the same selective-update rule should work with automatically detected change factors (e.g., labeling which input statistics shifted between consecutive tasks), replacing the manual labels the paper's Limitations section says are currently required.
  • My inference: the paper's experiments are consistent with contexts acting as low-dimensional control modes, so one testable implication is that learned robot-context vectors should cluster by physical property (mass, Kp, delay) in a way that survives dimensionality reduction; the paper does not report such an analysis.
  • My inference: a boundary condition worth probing is factor interaction—if terrain difficulty changes how much payload matters, independent vectors may not compose; testing on tasks with multiplicative factor interactions would show whether the method generalizes beyond additive factor combinations.
  • My inference: the real-world 80% result depends on selecting complementary contexts (sim stair terrain + real flat robot). The paper's dynamics-level wrong-combination results suggest the policy would degrade if both contexts came from the same flat real dataset; a deployment-level ablation replacing the sim terrain context with a real flat terrain context would make the orthogonality claim direct
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes DMCM, a CAVIA-style meta-learning algorithm that maintains K factor-specific context vectors rather than a single unified context. During meta-training, tasks are sampled conditionally so that consecutive tasks differ in exactly one labeled factor, and the inner loop updates only the context vector associated with that factor. An optional recombination loop trains the model to use context vectors adapted separately, enabling zero-shot recombination. The method is evaluated on sine regression (OOD robustness, zero-shot recombination) and on a Go1 quadruped pipeline in which contexts are first extracted by a dynamics model and then used to condition an RL policy. The authors report improved OOD robustness over MAML/CAVIA/ANIL in sine tasks and successful real-world stair climbing with 20 seconds of flat-terrain data by combining a simulated terrain context with a real-world robot-property context.

Significance. If the central claim holds, DMCM offers an interpretable, factor-aligned alternative to monolithic task embeddings in meta-learning and a practical route to sim-to-real context reuse in legged locomotion. The real-world deployment is ambitious and the context-sharing idea is genuinely useful. However, the current evidence does not isolate the proposed selective-update mechanism from the conditional task-sampling curriculum, and the test-time protocol for unknown factor changes is underspecified. These gaps are load-bearing for the paper's main claims. The authors are also candid about manual context labeling and other limitations, which is a strength.

major comments (3)
  1. [§3.2, Alg. 1 (lines 8 and 12), Figs. 3–4, App. A.2] The OOD robustness comparison does not include a CAVIA baseline trained on the same conditional chain p(Ti | Ti-1). DMCM is trained on a specially organized sequence in which consecutive tasks differ in exactly one labeled factor, while CAVIA is trained on i.i.d. tasks. The gains in Figs. 3–4 could therefore be due to the task curriculum itself rather than to the selective update of only phi_s. This is not a purely hypothetical concern: in the 5-shot results (App. A.2), DMCM and CAVIA perform similarly under exclusion, and in the mislabeling experiment (Table 4) both methods degrade comparably. To support the mechanistic claim, add ablations with (i) CAVIA trained on the same conditional chain, and (ii) a multi-context variant with the same chain but updating all context vectors, with matched context-parameter counts and compute.
  2. [§3.5, §4.1.1] The test-time adaptation protocol is underspecified. In sine OOD evaluation, test tasks are sampled with both amplitude and phase varying, so the model is not told which factor 'changed'. 'Sequential adaptation is done from context c1 to cK' does not explain how the relevant factor is chosen. If only one context is updated, the method needs an unstated label or inference mechanism; if all contexts are updated, DMCM's selective mechanism is not exercised at evaluation. Please state the exact evaluation algorithm (which contexts are updated, with what data, and for how many steps) and discuss how the method behaves when multiple factors change simultaneously.
  3. [§4.2.2, Table 3 (main text)] The headline real-world result is based on 5 trials per condition, so the 80% vs. 40% success rates for Multi-DMCM and Vanilla are not statistically distinguishable. In addition, the single-CAVIA policy is evaluated separately with each context dataset, but no baseline is provided for a unified context adapted on a combined dataset that contains both the simulated stair information and the real flat-terrain payload information. Please increase the number of trials, report confidence intervals or individual trial outcomes, and include a combined-context CAVIA baseline where feasible.
minor comments (6)
  1. [Alg. 1] The pseudocode specifies initialization of only phi_s but does not state how the non-selected context vectors are carried from the previous task. Please clarify the bookkeeping, especially in the basic (non-recombination) loop.
  2. [App. A.2] The 5-shot result, where DMCM loses its OOD advantage, is only in the appendix and is not discussed in the main text. This qualification should appear wherever the 10-shot OOD advantage is claimed.
  3. [Table 4, App. A.4] DMCM's clean-data loss (0.0225) is notably worse than CAVIA's (0.0135), and both methods degrade by similar proportions under mislabeling. This should be discussed as evidence about sensitivity to label noise, not as a robustness success.
  4. [General] There are duplicate table numbers: the main-text Table 3 and the Appendix Table 3 both exist. Please renumber.
  5. [§4.1.1] The phrase 'identical conditions (10 inner steps, same learning rate)' is misleading because DMCM also uses warm-up tasks and a conditional sampling chain. Please specify exactly which hyperparameters are matched and which are not.
  6. [Fig. 2] The diagram for K=2 is very small and does not convey the sequential adaptation and recombination procedure. A larger, annotated figure would help readers.

Circularity Check

1 steps flagged

Disentanglement is imposed by the labeled task curriculum and selective update rule, but the OOD and real-robot results are external; partial circularity in the central mechanism claim.

specific steps
  1. self definitional [§3.2 Inner Loop Training; Algorithm 1 lines 8, 12, 20; §3.1 Eq. (4); §6 Limitations]
    "At each step, only the context vector φsi corresponding to the changing factor cs is updated, while the others remain unchanged. Specifically, task Ci is chosen to differ from Ci−1 only in cs, with the other contexts identical. ... Sample Ti ∼ p(Ti | Ti−1); ... φsi ← φsi − α∇φs ..."

    The claimed ‘disentanglement’ of context vectors is not derived or discovered; it is hard-wired by the training procedure. Algorithm 1 line 8 supplies the label s of the changing factor through the conditional task distribution p(Ti|Ti−1), and line 12 updates exactly that φs. Thus φs is forced to align with factor cs by construction, and the statement ‘the context vector corresponding to the changing factor is updated’ is a restatement of the update rule plus hand-defined labels, not an empirical finding. The paper’s own limitation (‘relies on manually defined, meaningful context labels’) concedes that the factor decomposition is an input. The OOD robustness and real-robot successes are externally measured and not circular, but the mechanistic attribution ‘robustness stems from disentangli

full rationale

The only substantial circularity is in the claim that DMCM ‘learns’ disentangled context vectors: the factor identity s is supplied by the conditional task sampler and the same s selects the updated vector, so the alignment of φs with factor cs is true by construction. This is a self-definitional property of the algorithm, explicitly acknowledged in the Limitations. However, the paper does not rely on self-citations, and its main empirical contributions—sine OOD loss, sim-to-real stair climbing with 20 s of flat data, payload/gain robustness—are external benchmarks with real evaluation, not quantities fitted from the same labels. The missing CAVIA-with-same-curriculum ablation is a confound for the mechanism, not a circularity. Overall score 4 reflects a central mechanistic claim that is partly by construction while the headline OOD results retain independent content.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The method's contribution rests on hand-selected factors and context dimensions, a conditional task-sampling scheme specific to this paper, and the assumption that dynamics-model contexts survive transfer into RL. There are no new physical entities, but the factor-alignment capability is partly engineered through the data schedule rather than discovered.

free parameters (5)
  • Number of contexts K (per factor) = 2 for sine/locomotion, 3 for the three-factor sine test
    Set equal to the number of manually labeled factors; Appendix B.1 shows that a redundant 4-context configuration destabilizes training, so K is a sensitive design choice.
  • Context vector dimension per factor = 3 (sine), 20 (dynamics)
    Chosen by hand; Appendix B.4 shows 1 parameter fails and 6 parameters destabilize the 2-context model, so the dimension affects the claimed disentanglement behavior.
  • Warm-up tasks B = 10 or 20
    Hand-chosen; the outer loop only begins after B warm-up tasks, and the warm-up length influences how much context information accumulates before shared parameters are updated.
  • Adaptation depth Sadapt = 3 to 10 depending on task
    Hand-chosen per task family; the paper states that a too-low Sadapt prevents full adaptation even when the model has capacity.
  • Inner/outer learning rates, decay, meta-batch size = Inner LR 0.1 or 0.05 with 0.92 decay; meta LR 0.001 or 0.00033; batch 25 or 45
    Standard tuned hyperparameters; the paper adjusts them between experiments and lists them as matched across baselines.
axioms (4)
  • domain assumption Task labels (which factor changes) are known and correct
    Section 3.1 requires an 'additional task-labeling step'; the Limitations admit the method 'relies on manually defined, meaningful context labels with sufficient variation.'
  • ad hoc to paper Consecutive training tasks can be sampled conditionally so that only one factor varies between them
    Algorithm 1, line 8 samples Ti ~ p(Ti | Ti-1); this conditional sequencing is what makes the selective inner-loop update meaningful and is a restriction on the task distribution, not a general property of meta-learning benchmarks.
  • domain assumption The learned dynamics-model context vectors remain valid when transferred as priors into the RL policy
    Section 4.2 (Fig. 6) assumes context vectors extracted from a dynamics model trained on naive-policy data can be concatenated into a policy trained by PPO; the transfer is tested empirically but is a structural premise of the pipeline.
  • domain assumption Gradient-based inner-loop optimization on the selected context vector converges to a factor-aligned representation
    Section 3.2 and Algorithm 1 presume that updating only phi_s on tasks that differ only in factor s yields a vector specifically encoding that factor; no proof is given beyond empirical context-use tests.
invented entities (1)
  • Factor-specific context vectors (c1..cK) no independent evidence
    purpose: Represent and share task factors separately, enabling selective adaptation and cross-task reuse
    The paper provides internal and real-world evidence that contexts encode factors (Figs. 7, 18; App. D.6), but the vectors are learned latents with no externally predicted value, and the factorization itself is imposed by manual labels and data sequencing.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Disentangled Multi-Context Meta-Learning: Unlocking robust and Generalized Task Learning." pith.science (2026). https://pith.science/paper/PQDX4TSI

@misc{pith2026250901297,
  author       = {Pith},
  title        = {Pith review of: Disentangled Multi-Context Meta-Learning: Unlocking robust and Generalized Task Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PQDX4TSI}},
  note         = {Machine review of arXiv:2509.01297}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

In meta-learning and its downstream tasks, many methods rely on implicit adaptation to task variations, where multiple factors are mixed together in a single entangled representation. This makes it difficult to interpret which factors drive performance and can hinder generalization. In this work, we introduce a disentangled multi-context meta-learning framework that explicitly assigns each task factor to a distinct context vector. By decoupling these variations, our approach improves robustness through deeper task understanding and enhances generalization by enabling context vector sharing across tasks with shared factors. We evaluate our approach in two domains. First, on a sinusoidal regression task, our model outperforms baselines on out-of-distribution tasks and generalizes to unseen sine functions by sharing context vectors associated with shared amplitudes or phase shifts. Second, in a quadruped robot locomotion task, we disentangle the robot-specific properties and the characteristics of the terrain in the robot dynamics model. By transferring disentangled context vectors acquired from the dynamics model into reinforcement learning, the resulting policy achieves improved robustness under out-of-distribution conditions, surpassing the baselines that rely on a single unified context. Furthermore, by effectively sharing context, our model enables successful sim-to-real policy transfer to challenging terrains with out-of-distribution robot-specific properties, using just 20 seconds of real data from flat terrain, a result not achievable with single-task adaptation.

Figures

Figures reproduced from arXiv: 2509.01297 by Jun-Gill Kang, Seongil Hong, Seonsoo Kim, Taehong Kim.

Figure 1
Figure 1. Figure 1: Basic Concept of Disentangled Multi-Context Meta-Learning. Adaptation is illustrated for both the sine regression and robot dynamics tasks. The model adapts by disentangling task-specific factors into separate context vectors. These learned contexts can be reused across tasks with overlapping factors, enabling generalization. *Equal contribution. 9th Conference on Robot Learning (CoRL 2025), Seoul, Korea. … view at source ↗
Figure 2
Figure 2. Figure 2: Simple diagram of DMCM for K=2 case To support zero-shot generalization when context vectors are shared, DMCM can optionally employ a recombination loop. In this procedure, the model parameters θ are updated using context vectors that were not adapted together during the inner loop. This encourages the model to work effectively with independently adapted context vectors. The de￾tailed procedure and its eff… view at source ↗
Figure 4
Figure 4. Figure 4: Average loss comparisons with range exclu￾sion. (a) At 2000 meta gradients (near convergence for CAVIA/DMCM on full range), (b) At 4000 meta gra￾dients (CAVIA/DMCM fully converged on full range). The shaded region shows standard deviation. 30 ran￾domly selected ranges were used for evaluation. We partition the amplitude and phase shift into five intervals, creating 25 range combinations. We then simulate o… view at source ↗
Figure 5
Figure 5. Figure 5: Zero-shot prediction with disentangled context vectors. Dotted lines (green, blue, orange) show predictions after adaptation to the displayed points, while the solid line denotes the ground truth. Red dotted lines indicate predictions using only the shared context vectors without adaptation [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Learning procedure at Quadrupedal Robot Locomotion Task 4.2 Quadrupedal Robot Locomotion Task In this section, we first train the robot dynamics model using both CAVIA and DMCM, then extract their latent features—context parameters (CAVIA) and context vectors (DMCM)—across diverse simulation conditions. These extracted contexts are subsequently transferred as prior knowledge in RL, illustrated in [PITH_FU… view at source ↗
Figure 7
Figure 7. Figure 7: Each subplot shows the average loss for each dataset using 16 context vector sets derived from four real-world datasets. Each set includes a terrain (Flat (F) or Wavy (W)) and a robot property context (Additional Mass (M+) or No Additional Mass (MX)). Results are averaged over 10 trials, with adaptation and test sets randomly sampled from 20 seconds within the 40 seconds of data. Error bars indicate the st… view at source ↗
Figure 8
Figure 8. Figure 8: Additional deployment results showing the multi￾DMCM policy on wavy terrain (left) and stair climbing (right) un￾der asymmetric payloads with water bottles and a lidar sensor. Using these contexts, the multi￾DMCM policy achieves the best performance in climbing steep stairs at OOD conditions, leveraging its inherent robustness, whereas the single-CAVIA policy fails com￾pletely. Moreover, by sharing context… view at source ↗
Figure 9
Figure 9. Figure 9: Loss curves for (a) No exclusion, (b) average loss comparison with range exclusions at 2000 meta￾gradients, and (c) 4000 meta-gradients Unlike the results in the 10-shot cases shown in [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: Example of an incorrect zero-shot pre￾diction when the phase-shift context is derived from a low-amplitude case (green). The zero-shot prediction (red dot) deviates significantly from the ground truth (red solid line) As shown in [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Additional zero-shot results from DMCM with two contexts. Zero-shot predictions (red dots) are obtained through shared context vectors. Tasks used for adapting amplitude and phase-shift contexts are shown in blue and green, respectively. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Training curves of N-context models under (a) full-range and (b) OOD evaluation. Shaded areas represent confidence intervals: 95% for (a) and 0.25 standard deviation for (b). The two-context case without the phase-shift factor is excluded in (b) due to extremely poor performance. In the full-range setting, all models perform well unless a critical variation is excluded from the contexts (the two-context m… view at source ↗
Figure 14
Figure 14. Figure 14: Additional zero-shot results from DMCM with three contexts. Zero-shot predictions (red dots) are obtained through shared context vectors. Tasks used for adapting amplitude, phase-shift, and y-shift contexts are shown in blue, green, and orange, respectively. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Training curves for the 2-context model (one entangled amplitude–phase context and one y-shift context) with different numbers of parameters for one context vector. Shaded regions indicate the 95% confi￾dence interval. We evaluate the effect of varying the number of parameters in each context vector using a 2-context model (one entangled amplitude–phase context and one y-shift context). As shown in [PITH… view at source ↗
Figure 16
Figure 16. Figure 16: Dynamics model loss for CAVIA, DMCM with self-adaptation, and DMCM with re￾combination (zero-shot with shared context vectors). Shaded region indicates 95% confidence interval. Evaluation is performed on 500 tasks. 0 10000 20000 30000 40000 50000 60000 Meta Gradient Steps 0.030 0.025 0.020 0.015 0.010 0.005 0.000 0.005 Loss RecombinationLoss - RandomContextLoss [PITH_FULL_IMAGE:figures/full_fig_p017_16.png] view at source ↗
Figure 18
Figure 18. Figure 18: Each subplot shows the average loss for each dataset using 16 context vector sets derived from four datasets. Each set includes a terrain context (Flat (F) or Stair (St)) and a robot property context (Sim (Si) or Real (R)). Results are averaged over 10 trials, with adaptation and test sets randomly sampled from 20 seconds within the 40 seconds of data. Error bars indicate the standard deviation. As shown … view at source ↗
Figure 19
Figure 19. Figure 19: Terrains used for evaluation 21 [PITH_FULL_IMAGE:figures/full_fig_p021_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Go1 robot with water bottles attached at three legs Using context vectors obtained from previously collected data, the multi-DMCM policy successfully navigates both stair and wavy terrains. For the stair task, the terrain context is taken from simulation data of an 18cm stair climb, and the robot-specific context is taken from real-world flat-terrain data with the water-bottle payload. For the wavy terrai… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Autonomous Aerial Manipulation via Contextual Contrastive Meta Reinforcement Learning

    cs.LG 2026-06 unverdicted novelty 5.0

    Aco2 trains a quadrotor policy in simulation that adapts to diverse payload dynamics via latent context encoding and contrastive structuring, enabling zero-shot real-world deployment for autonomous aerial delivery.

Reference graph

Works this paper leans on

42 extracted references · 34 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  1. [1]

    Thrun and L

    S. Thrun and L. Pratt. Learning to learn: Introduction and overview. In Learning to learn , pages 3–17. Springer, 1998. 2

    work page 1998

  2. [2]

    C. Finn, T. Yu, T. Zhang, P. Abbeel, and S. Levine. One-shot visual imitation learning via meta-learning. In Conference on robot learning, pages 357–368. PMLR, 2017. 2

    work page 2017

  3. [3]

    Kaushik, T

    R. Kaushik, T. Anne, and J.-B. Mouret. Fast online adaptation in robotics through meta- learning embeddings of simulated priors. In 2020 IEEE/RSJ International Conference on In- telligent Robots and Systems (IROS), pages 5269–5276. IEEE, 2020

    work page 2020

  4. [4]

    Nagabandi, I

    A. Nagabandi, I. Clavera, S. Liu, R. S. Fearing, P. Abbeel, S. Levine, and C. Finn. Learning to adapt in dynamic, real-world environments through meta-reinforcement learning. arXiv preprint arXiv:1803.11347, 2018

    Pith/arXiv arXiv 2018

  5. [5]

    O’Connell, G

    M. O’Connell, G. Shi, X. Shi, K. Azizzadenesheli, A. Anandkumar, Y . Yue, and S.-J. Chung. Neural-fly enables rapid learning for agile flight in strong winds. Science Robotics, 7(66): eabm6597, 2022. 2

    work page 2022

  6. [6]

    A. A. Rusu, D. Rao, J. Sygnowski, O. Vinyals, R. Pascanu, S. Osindero, and R. Hadsell. Meta- learning with latent embedding optimization. arXiv preprint arXiv:1807.05960, 2018. 2

    Pith/arXiv arXiv 2018

  7. [7]

    Zintgraf, K

    L. Zintgraf, K. Shiarli, V . Kurin, K. Hofmann, and S. Whiteson. Fast context adaptation via meta-learning. In International conference on machine learning , pages 7693–7702. PMLR,

    work page

  8. [8]

    Raghu, M

    A. Raghu, M. Raghu, S. Bengio, and O. Vinyals. Rapid learning or feature reuse? towards understanding the effectiveness of maml. arXiv preprint arXiv:1909.09157, 2019. 2, 4

    Pith/arXiv arXiv 1909

  9. [9]

    Z. Li, F. Zhou, F. Chen, and H. Li. Meta-sgd: Learning to learn quickly for few-shot learning. arXiv preprint arXiv:1707.09835, 2017

    Pith/arXiv arXiv 2017

  10. [10]

    Nichol and J

    A. Nichol and J. Schulman. Reptile: a scalable metalearning algorithm. arXiv preprint arXiv:1803.02999, 2(3):4, 2018. 2 9

    Pith/arXiv arXiv 2018

  11. [11]

    C. Finn, P. Abbeel, and S. Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2, 4

    work page 2017

  12. [12]

    F. Alet, T. Lozano-P´erez, and L. P. Kaelbling. Modular meta-learning. In Conference on robot learning, pages 856–868. PMLR, 2018. 2

    work page 2018

  13. [13]

    Rakelly, A

    K. Rakelly, A. Zhou, C. Finn, S. Levine, and D. Quillen. Efficient off-policy meta- reinforcement learning via probabilistic context variables. In International conference on ma- chine learning, pages 5331–5340. PMLR, 2019. 2

    work page 2019

  14. [14]

    H. Yao, Y . Wei, J. Huang, and Z. Li. Hierarchically structured meta-learning. In International conference on machine learning, pages 7045–7054. PMLR, 2019. 2

    work page 2019

  15. [15]

    Vuorio, S.-H

    R. Vuorio, S.-H. Sun, H. Hu, and J. J. Lim. Multimodal model-agnostic meta-learning via task-aware modulation. Advances in neural information processing systems, 32, 2019. 2

    work page 2019

  16. [16]

    M. W. Gondal, M. Wuthrich, D. Miladinovic, F. Locatello, M. Breidt, V . V olchkov, J. Akpo, O. Bachem, B. Sch ¨olkopf, and S. Bauer. On the transfer of inductive bias from simulation to the real world: a new disentanglement dataset. Advances in Neural Information Processing Systems, 32, 2019. 2

    work page 2019

  17. [17]

    K. Hsu, J. I. Hamid, K. Burns, C. Finn, and J. Wu. Tripod: Three complementary inductive biases for disentangled representation learning. arXiv preprint arXiv:2404.10282, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  18. [18]

    M. F. Mathieu, J. J. Zhao, J. Zhao, A. Ramesh, P. Sprechmann, and Y . LeCun. Disentan- gling factors of variation in deep representation using adversarial training. Advances in neural information processing systems, 29, 2016

    work page 2016

  19. [19]

    X. Chen, Y . Duan, R. Houthooft, J. Schulman, I. Sutskever, and P. Abbeel. Infogan: Inter- pretable representation learning by information maximizing generative adversarial nets. Ad- vances in neural information processing systems, 29, 2016

    work page 2016

  20. [20]

    Bengio, L

    Y . Bengio, L. Yao, G. Alain, and P. Vincent. Generalized denoising auto-encoders as generative models. Advances in neural information processing systems, 26, 2013

    work page 2013

  21. [21]

    Higgins, L

    I. Higgins, L. Matthey, A. Pal, C. Burgess, X. Glorot, M. Botvinick, S. Mohamed, and A. Ler- chner. beta-vae: Learning basic visual concepts with a constrained variational framework. In International conference on learning representations, 2017

    work page 2017

  22. [22]

    Higgins, A

    I. Higgins, A. Pal, A. Rusu, L. Matthey, C. Burgess, A. Pritzel, M. Botvinick, C. Blundell, and A. Lerchner. Darla: Improving zero-shot transfer in reinforcement learning. In International conference on machine learning, pages 1480–1490. PMLR, 2017. 2

    work page 2017

  23. [23]

    L. P. Kaelbling, M. L. Littman, and A. R. Cassandra. Planning and acting in partially observ- able stochastic domains. Artificial Intelligence, 101(1-2):99–134, 1998. 3

    work page 1998

  24. [24]

    Kumar, Z

    A. Kumar, Z. Fu, D. Pathak, and J. Malik. Rma: Rapid motor adaptation for legged robots. arXiv preprint arXiv:2107.04034, 2021. 3

    Pith/arXiv arXiv 2021

  25. [25]

    I. M. A. Nahrendra, B. Yu, and H. Myung. Dreamwaq: Learning robust quadrupedal lo- comotion with implicit terrain imagination via deep reinforcement learning. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 5078–5084. IEEE, 2023. 3

    work page 2023

  26. [26]

    P. Wu, W. Xie, J. Cao, H. Lai, and W. Zhang. Loopsr: Looping sim-and-real for lifelong policy adaptation of legged robots. arXiv preprint arXiv:2409.17992, 2024. 3

    work page arXiv 2024

  27. [27]

    Z. Xiao, X. Zhang, X. Zhou, and Q. Zhang. Pa-loco: Learning perturbation-adaptive locomo- tion for quadruped robots. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 9110–9115. IEEE, 2024. 3

    work page 2024

  28. [28]

    Makoviychuk, L

    V . Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, et al. Isaac gym: High performance gpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470, 2021. 6, 16, 18 10

    Pith/arXiv arXiv 2021

  29. [29]

    Hwangbo, J

    J. Hwangbo, J. Lee, and M. Hutter. Per-contact iteration method for solving contact dynamics. IEEE Robotics and Automation Letters, 3(2):895–902, 2018. 18

    work page 2018

  30. [30]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 18

    Pith/arXiv arXiv 2017

  31. [31]

    G. Ji, J. Mun, H. Kim, and J. Hwangbo. Concurrent training of a control policy and a state estimator for dynamic and robust legged locomotion. IEEE Robotics and Automation Letters, 7(2):4630–4637, 2022. 19

    work page 2022

  32. [32]

    Rudin, D

    N. Rudin, D. Hoeller, P. Reist, and M. Hutter. Learning to walk in minutes using massively parallel deep reinforcement learning. Conference on Robot Learning (PMLR), pages 91–100,

    work page

  33. [33]

    10% mislabeling

    G. B. Margolis and P. Agrawal. Walk these ways: Tuning robot control for generalization with multiplicity of behavior. Conference on Robot Learning (PMLR), pages 22–31, 2023. 19 11 Appendix A Sine Task A.1 Hyperparameter and Experiment Settings To ensure a fair comparison across MAML, ANIL, CA VIA, and DMCM, we align the hyperparam- eters as closely as po...

    work page 2023

  34. [34]

    Detailed input and hyperparameters are in Table 7, 8

    to enable robust sim-to-real transfer. Detailed input and hyperparameters are in Table 7, 8. Observation Type Input Dim. Proprioception Linear body velocity estimation 3Angular body velocity 3Body height estimation 1Foot height estimation 4Contact probability estimation 4Command 3Projected gravity vector 3Action 12Joint position 12Joint velocity 12Action ...

    work page 2000

  35. [35]

    Flat terrain with 1.5kg payload

    work page

  36. [36]

    Climbing stairs (17cm depth)

    work page

  37. [37]

    Wavy terrain with 1.5kg payload

    work page

  38. [38]

    Flat terrain with 1.5kg payload and water bottle on legs

    work page

  39. [39]

    These datasets are used for adapting the CA VIA and DMCM dynamics models

    Flat terrain with 16N · m Kp (low gain) and 1.5kg payload Data collection on stairs with a 1.5 kg payload was not possible, as the naive policy was unable to successfully complete the task. These datasets are used for adapting the CA VIA and DMCM dynamics models. The adapted contexts are then applied in the single-CA VIA and multi-DMCM policies for real-w...

    work page

  40. [40]

    High stairs down (with velocity limits)

    work page

  41. [41]

    High stairs up (with velocity limits)

    work page

  42. [42]

    High stairs up (with velocity limits and low friction) Velocity Limits: For high stairs, the linear command x directional velocity ranges from 0.4 m/s to 0.6 m/s, y directional velocity as 0, and yaw rate from −0.6 rad/s to 0.6 rad/s. Except for high stairs, the linear command x directional velocity ranges from −1.0 m/s to 1.0 m/s, y directional velocity ...

    work page 2000

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.