Pith. sign in

REVIEW 2 major objections 5 minor 34 references

Dynamic Mixture of Progressive Parameter-Efficient Expert Library for Lifelong Robot Learning

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that by growing a library of frozen low-rank experts and replaying the router's decisions, a frozen pretrained robot policy can keep learning new manipulation tasks throughout its lifetime without an oracle task…

desk verdict Solid empirical paper on lifelong robot learning; the product-of-sums formula in Eq. (7) is a real interpretive gap, but the results stand. read the letter →

arxiv 2506.05985 v3 pith:6EFKIVR2 submitted 2025-06-06 cs.LG cs.RO

Yuheng Lei , Sitong Mao , Shunbo Zhou , Hongyuan Zhang , Xuelong Li , Ping Luo This is my paper
classification cs.LGcs.RO
keywords lifelonglearningrobotmanipulationparameter-efficientfine-tuninglow-rankadaptationmixtureofexpertscatastrophicforgettingforwardtransfer
verification ladder T0 review T1 audit T2 compute T3 formal

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 asks whether a frozen pretrained robot policy can keep learning new manipulation tasks across a lifetime without forgetting old ones and without being told which task it faces. It answers yes with a design that grows a library of small low-rank adapters, one per encountered task, and lets a lightweight router mix the frozen adapters on the fly from the current visual, language, and proprioceptive context. Because each new task adds only a new adapter and the router is the only cross-task component, the paper adds expert coefficient replay—storing a small buffer of router input–output pairs—to keep the router from drifting away from old tasks. On the LIBERO lifelong benchmark, the method reports higher forward transfer and lower negative backward transfer than replay, regularization, and adapter-retrieval baselines while training only about 1.2 million parameters and storing only low-dimensional coefficient replay data.

What carries the argument

The central mechanism is the low-rank expert library paired with a context-conditioned router. Each selected linear layer of the frozen policy holds a growing library of LoRA-style experts $\{A_j, B_j, b_j\}$; the router $\mathcal{R}$ maps the mean-pooled frozen context embedding—visual, textual, and proprioceptive—to per-module coefficient vectors, retains the top-$k$ entries, and synthesizes the effective layer weights as $W_0 + \sum_j c_j A_j B_j$ plus bias terms. New experts are zero-initialized except for an orthonormalized $A$ matrix, so they begin by copying prior behavior and can specialize; expert coefficient replay stores $(r, c)$ pairs and adds an MSE term that keeps the router faithful to previously learned tasks.

What would settle it

Run DMPEL on a longer concatenated task sequence than the 30-task one and measure success on early tasks; a monotone decline that is not recovered by enlarging the router, or an ablation in which expert coefficient replay is removed but negative backward transfer stays near zero, would contradict the paper's explanation of how forgetting is controlled.

Watch

Extended reading notes

Core claim

DMPEL's central claim is that catastrophic forgetting in lifelong parameter-efficient fine-tuning can be moved from the policy parameters into a small router and then contained by replaying the router's own decisions. The paper shows that when each arriving task contributes a frozen low-rank expert to a shared library, a top-k router can synthesize an effective per-task policy from the frozen experts and the newly added expert, using only mean-pooled frozen CLIP embeddings of the current observation as input. The same modularity lets the method store a small buffer of context-embedding and expert-coefficient pairs from past tasks and regularize the router with a mean-squared-error loss, so old tasks are retrieved accurately without storing or replaying demonstration trajectories. The paper's experiments across the four LIBERO suites and a 30-task concatenated sequence report that this scheme achieves better forward transfer and near-zero forgetting than full-fine-tuning replay and regularization baselines, and than LoRA methods that rely on oracle task IDs, using 1.2M trainable parameters.

Load-bearing premise

The load-bearing premise is that the small two-hidden-layer router can learn and keep a correct mapping from frozen context embeddings to expert-mixing coefficients for every task encountered, including long sequences; the paper itself notes that as the number of distinct tasks grows, the router's plasticity may be exhausted.

Editorial extensions

If this is right

  • A frozen pretrained policy can be continually adapted with only about 0.7% of its parameters trainable, making lifelong adaptation affordable for large vision-language-action models.
  • Because replay stores only context embeddings and expert coefficients, storage is roughly 6.7% of full experience replay and the computation avoids running the whole policy on old trajectories.
  • Underactivated experts can be pruned without hurting success, keeping the expert library compact over long task sequences.
  • The method removes the need for an oracle task identifier at test time, so deployment does not require knowing which task is being solved.
  • Larger router capacity partially recovers performance on very long sequences, suggesting that router size, not expert count, is the main scaling bottleneck.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same coefficient-replay trick should transfer to other modular parameter-efficient methods—prompts, adapters, or task vectors—because it only requires an input-conditioned coefficient vector and a router to replay.
  • The router's input is mean-pooled frozen CLIP embeddings; a distribution shift such as a new embodiment or real-world camera images may require router recalibration, a question the paper leaves open.
  • A testable extension would be replacing the fixed MLP router with a growing or recurrent router and measuring whether the gradual decline on long sequences flattens without sacrificing replay efficiency.
  • The paper's expert-activation visualizations imply that ordering effects matter: the first task's action-head expert is heavily reused, so choosing which tasks arrive early could be as influential as the router architecture.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes DMPEL, a lifelong learning method for vision-language robot manipulation. A frozen pretrained policy is adapted through a growing library of low-rank LoRA experts, combined by a lightweight router that maps mean-pooled CLIP context embeddings to per-module expert coefficients. Expert coefficient replay stores context-coefficient pairs from previous tasks and regularizes the router via MSE to prevent forgetting. The method is evaluated on four LIBERO suites (10 tasks each), a 30-task concatenated sequence, and a cross-domain Robomimic sequence, with ablations on top-k, rank, replay ratio, task order, demos, and router design. The paper reports that DMPEL outperforms baselines on forward transfer and negative backward transfer while using 1.2M trainable parameters and small replay storage.

Significance. If the reported results hold, DMPEL is a practically attractive approach to lifelong VLA adaptation: it removes the oracle task-ID assumption of methods like TAIL, achieves forward transfer through dynamic composition of prior knowledge, and introduces a cheap replay surrogate (coefficient replay) that mitigates forgetting. The empirical side is thorough: three seeds, four benchmark suites, many ablations (top-k, rank, CR ratio, task order, demos, router size), and a 30-task stress test. The authors also explicitly discuss limitations (router capacity, sim2real) in Appendices B.4 and C.1, which increases confidence in the claims. The main weakness is conceptual: the parameter synthesis in Eq. (7) is not a standard mixture of experts, and the paper does not examine whether its performance story depends on that choice.

major comments (2)
  1. [4.2, Eq. (7)] The paper's expert-reuse narrative is not supported by the parameter synthesis formula. Eq. (7) defines the adapted linear layer as x(Σ_j c_j A_j)(Σ_j c_j B_j), which is the product of two coefficient-weighted sums, not the standard mixture of LoRA experts ΔW = Σ_j c_j A_j B_j. With top-3 routing, the effective update contains cross terms c_i c_j A_i B_j for i≠j, meaning the router does not retrieve or combine frozen experts as individual rank-r updates; it produces one rank-r update whose factors are linear combinations of the active A's and B's. This undermines the interpretation of Figures 6 and 8 (expert activation/reuse) and of Eq. (8) (coefficient replay as 'retrieving frozen experts'), because a stored coefficient vector does not have the same semantics under the two formulas. The paper should (a) justify the product-of-sums design or replace it with the standard sum-of-products, and (b) ablate Eq. (7) against ΔW = Σ c_j A_j B_j and against a single LoRA with matched active rank to show whether the reported FWT/NBT gains depend on the product form.
  2. [Algorithm 1, line 29] When a low-rank expert is pruned, the router's output dimension is reduced by removing the corresponding dimension. However, the coefficient replay buffer B_CR stores coefficient vectors from previous tasks that include the pruned dimension. Eq. (8) then computes MSE between R(r_i) (with reduced output dimension) and c_{i,old} (with the original dimension), which is undefined. The paper does not describe how stored coefficients are truncated or remapped after pruning. If pruning is not used in the main experiments, this should be stated explicitly; if it is used, the replay mechanism needs a concrete handling of pruned dimensions.
minor comments (5)
  1. [3.2 and Tables 4, 6, 8] The metric defined in Eq. (3) is called NBT in the main text, and the text says 'lower NBT indicates less catastrophic forgetting,' but the appendix tables label the same column as BWT. Please unify the terminology to avoid confusion.
  2. [4.2, Eq. (6) and Eq. (7)] The symbol c_t is used both for the full coefficient set in R^{M×K} and for the per-sub-module vectors c^⋄_t in R^K. Please clarify the notation in the walk-through after Eq. (7).
  3. [5.2] The statement 'only 1.2M trainable parameters (less than 0.7% of the policy)' refers to the number of parameters being optimized at a single task, not the total expert library, which grows with the number of tasks (shown in Table 7 as about 10.7–11.7M after 30 tasks). Please state this distinction explicitly when making the efficiency claim.
  4. [Appendix A.2] The claim 'we observed convergence typically within 10 epochs' is asserted without immediate evidence; the learning curves in Figures 18–21 appear to support it, so consider referencing them at that sentence.
  5. [5.2, Figure 5b] The statement that coefficient replay has 'approx. 6.7%' storage overhead relative to full experience replay would be clearer if the calculation (embedding dimension, coefficient dimension, and replay ratio) were given in the text or caption.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DMPEL's claims rest on external benchmark evaluations, not on self-referential derivation.

full rationale

DMPEL is an empirical lifelong learning method, not a mathematical derivation whose conclusion is smuggled in through an assumption. The central claims—better forward transfer, lower catastrophic forgetting, and small trainable parameter counts—are evaluated on the LIBERO benchmark against external baselines (ER, EWC, PackNet, LOTUS, TAIL, L2M, IsCiL) using success-rate metrics computed in simulation. The paper's design components are defined operationally and then tested: the low-rank expert library, the router, and expert coefficient replay. Equation (7) defines the adapted layer as a product of coefficient-weighted sums, which is an unusual modeling choice and may weaken the 'mixture of experts' narrative, but it is not a circular argument: the reported FWT, NBT, and AUC numbers are not produced by Eq. (7) by construction; they come from environment rollouts. Equation (8) regularizes the router to reproduce its own stored coefficients, which is self-referential as a mechanism, but again not a hidden circularity, because the benchmark metrics are not defined in terms of the replay loss. The paper's few self-citations (AdaptFormer and pi-tuning, both involving co-author Ping Luo) appear in related work as examples of PEFT and are not load-bearing for the central claim; no uniqueness theorem or prior result is invoked to force the method's choice. Appendix C.1 candidly states that 'when the number of distinct tasks continue to increase, the plasticity of the router may be exhausted,' and Figure 16 shows a gradual decline on a 30-task sequence; this is a stated limitation, not evidence that the result is assumed. Overall, the paper is self-contained against external benchmarks and exhibits no significant circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim depends on standard imitation-learning assumptions (behavior cloning, pretrain-then-finetune), the sufficiency of the router's context embedding for expert selection, and the unvalidated product-of-sums expert synthesis. The free parameters are the hand-chosen hyperparameters that the final performance depends on: top-3, CR=5%, pruning threshold 0.01, output range [0,2], rank sizes, and synthesis interval. No new physical entities are introduced.

free parameters (6)
  • top-k = 3
    Number of experts activated at each step; selected via ablation in Figure 5a.
  • coefficient replay ratio = 5%
    Fraction of router input-output pairs stored per task; selected via ablation in Figure 5b.
  • pruning threshold xi = 0.01
    Experts with accumulated coefficient below this are pruned; chosen by hand in Section 4.2.
  • scaled sigmoid output range = [0,2]
    Coefficient bounds; set from intuition, midpoint 1 means unchanged LoRA (Section B.1).
  • LoRA rank = 8 for CLIP encoders, 16 for others
    Borrowed from TAIL (Liu et al., 2024b); ablation shows saturation at rank 32.
  • expert synthesis interval = 1
    Weights re-synthesized every timestep in main experiments; ablation shows trade-off in Figure 5c.
assumptions (4)
  • standard math Behavior cloning objective (negative log-likelihood of actions under a Gaussian Mixture head) is the training signal for all methods.
    Eq. (1); standard imitation learning setup.
  • domain assumption The policy pretrained on LIBERO-90 provides a useful starting point for downstream lifelong learning.
    Section 5.1; the pretrain-then-finetune paradigm is assumed, not derived.
  • domain assumption The router input r_t, formed by mean-pooling frozen CLIP embeddings and proprioceptive states, contains enough information to select the correct expert combination for each task and context.
    Section 4.2; if this fails, routing and therefore the whole method fails.
  • ad hoc to paper The dynamic parameter synthesis in Eq. (7), where LoRA updates are (sum c_j A_j)(sum c_j B_j), is a valid and useful way to combine experts.
    No theoretical justification; this product-of-sums form creates cross terms and is not ablated against the standard sum of c_j A_j B_j.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dynamic Mixture of Progressive Parameter-Efficient Expert Library for Lifelong Robot Learning." pith.science (2026). https://pith.science/paper/6EFKIVR2

@misc{pith2026250605985,
  author       = {Pith},
  title        = {Pith review of: Dynamic Mixture of Progressive Parameter-Efficient Expert Library for Lifelong Robot Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6EFKIVR2}},
  note         = {Machine review of arXiv:2506.05985}
}
read the original abstract

A generalist agent must continuously learn and adapt throughout its lifetime, achieving efficient forward transfer while minimizing catastrophic forgetting. Previous work within the dominant pretrain-then-finetune paradigm has explored parameter-efficient fine-tuning for single-task adaptation, effectively steering a frozen pretrained model with a small number of parameters. However, in the context of lifelong learning, these methods rely on the impractical assumption of a test-time task identifier and restrict knowledge sharing among isolated adapters. To address these limitations, we propose Dynamic Mixture of Progressive Parameter-Efficient Expert Library (DMPEL) for lifelong robot learning. DMPEL progressively builds a low-rank expert library and employs a lightweight router to dynamically combine experts into an end-to-end policy, enabling flexible and efficient lifelong forward transfer. Furthermore, by leveraging the modular structure of the fine-tuned parameters, we introduce expert coefficient replay, which guides the router to accurately retrieve frozen experts for previously encountered tasks. This technique mitigates forgetting while being significantly more storage- and computation-efficient than experience replay over the entire policy. Extensive experiments on the lifelong robot learning benchmark LIBERO demonstrate that our framework outperforms state-of-the-art lifelong learning methods in success rates during continual adaptation, while utilizing minimal trainable parameters and storage.

Figures

Figures reproduced from arXiv: 2506.05985 by the authors.

Figure 1
Figure 1. Existing lifelong learning methods: (a) Replay methods; (b) Regularization methods; (c) Ar [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed method DMPEL. Final SR = 1 κ Xκ K=1 sκ,K. (5) 4 Proposed Method In order to address the key challenges in lifelong robot learning, we propose Dynamic Mixture of Progressive Parameter-Efficient Expert Library (DMPEL). To enhance knowledge transfer, DMPEL incrementally builds a library of low-rank experts and uses a lightweight router to dynamically compose them into a unified policy based on … view at source ↗
Figure 3
Figure 3. Performance of different lifelong robot learning methods on the LIBERO benchmark. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (18 more)
Figure 5
Figure 5. Figure 5: Ablation studies. (a) Sparse activation of LoRA experts and the granularity of expert coefficients [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 4
Figure 4. Figure 4: Performance of the pre￾trained policy on LIBERO Ablation Studies. In Figure 5a, we conduct an ablation study across two dimensions: the sparse activation of LoRA experts in the library (using top-1 expert, using top-3 experts, and soft merging all experts) and the gran…
Figure 6
Figure 6. Figure 6: Analysis on expert coefficients on LIBERO-Object. (a) Visualization of the coefficients for each [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: t-SNE visualization of embeddings from the final block when evaluated on Task 2 from LIBERO-Object. t-SNE Analysis. We present t-SNE visualization in [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Visualization Analysis on Expert Activation in LIBERO-Spatial: (a)-(b) Task 4: pick up the black [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Ablation studies in the LIBERO-Goal suite. (a) Different task orders (b) Different rank sizes (c) [PITH_FULL_IMAGE:figures/full_fig_p023_9.png]
Figure 10
Figure 10. Figure 10: Ablation studies on router design in the LIBERO-Goal suite. [PITH_FULL_IMAGE:figures/full_fig_p023_10.png]
Figure 11
Figure 11. Figure 11: Visualization of the action trajectories from demonstrations of LIBERO-Object benchmark. [PITH_FULL_IMAGE:figures/full_fig_p025_11.png]
Figure 12
Figure 12. Figure 12: Visualization Analysis on LoRA Expert Similarities across Tasks. [PITH_FULL_IMAGE:figures/full_fig_p026_12.png]
Figure 13
Figure 13. Figure 13: Visualization of the action trajectories from demonstrations of LIBERO-Object benchmark. [PITH_FULL_IMAGE:figures/full_fig_p026_13.png]
Figure 14
Figure 14. Figure 14: Visualization Analysis on Expert Activation: (a)-(b) Task 2 from LIBERO-Goal: put the bowl on [PITH_FULL_IMAGE:figures/full_fig_p027_14.png]
Figure 15
Figure 15. Figure 15: Visualization Analysis on Expert Activation: Task 5 from LIBERO-Object: pick up the ketchup [PITH_FULL_IMAGE:figures/full_fig_p027_15.png]
Figure 16
Figure 16. Figure 16: Average success rate on encountered tasks in the ultra long task sequence concatenated with [PITH_FULL_IMAGE:figures/full_fig_p028_16.png]
Figure 17
Figure 17. Figure 17: Success rate on task 2, 5, 8, 13, 15, 18 in the ultra long task sequence. [PITH_FULL_IMAGE:figures/full_fig_p028_17.png]
Figure 18
Figure 18. Figure 18: Visualization of learning curves on the LIBERO-Goal benchmark. [PITH_FULL_IMAGE:figures/full_fig_p029_18.png]
Figure 19
Figure 19. Figure 19: Visualization of learning curves on the LIBERO-Spatial benchmark. [PITH_FULL_IMAGE:figures/full_fig_p030_19.png]
Figure 20
Figure 20. Figure 20: Visualization of learning curves on the LIBERO-Object benchmark. [PITH_FULL_IMAGE:figures/full_fig_p030_20.png]
Figure 21
Figure 21. Figure 21: Visualization of learning curves on the LIBERO-Long benchmark. [PITH_FULL_IMAGE:figures/full_fig_p031_21.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 21 canonical work pages

  1. [2]

    Similarly, in Figure 15, the vision encoder reuses previously learned expert but the fusion sub-module learns a new expert to effectively fuse features for manipulation

    In task 9 that also involves picking up a bowl but placing it down at a different destination, the vision encoder consistently activates expert 2 throughout the trajectory, while the temporal decoder primarily reuses expert 2 during the first half of the trajectory, likely due to a similar action pattern. Similarly, in Figure 15, the vision encoder reuses...

    work page 2020

  2. [3]

    Each task provides 50 successful expert demonstrations for imitation learning

    Although some tasks have similar descriptions, they are not identical due to differences in 20 Published in Transactions on Machine Learning Research (05/2026) Table 3: Tasks Instructions of LIBERO Task Suites Benchmark Suite Task Instructions open the middle drawer of the cabinet put the bowl on the stove put the wine bottle on top of the cabinet open th...

    work page 2026

  3. [4]

    The selected parameters are then fine-tuned and frozen to prevent forgetting

    • PackNet (Mallya & Lazebnik, 2018), anarchitecturalmethod that first trains and prunes the network to retain a fixed proportion (25% in our experiments) of the most important parameters. The selected parameters are then fine-tuned and frozen to prevent forgetting. • LifelOng knowledge Transfer Using Skills (LOTUS) (Wan et al.,

    work page 2018

  4. [5]

    (a) Different task orders (b) Different rank sizes (c) Different number of demonstrations

    Although there is only a slight decrease in success rate (e.g., 2%), this allows for adaptive scaling based on computational Table 4: Performance on LIBERO-Goal with Different Rank Size Method Rank Size FWT BWT AUC TAIL (with task ID) rank 0.54±0.03 0 0.67±0.06 TAIL (with task ID) rank×3 0.55±0.03 0 0.71±0.02 ISCiL rank 0.54±0.04 0.05±0.03 0.60±0.06 ISCiL...

    work page 2026

  5. [6]

    We present results using the policy ViT-T provided in the original LIBERO benchmark with two different sizes: the smaller model has 1.9M parameters (similar to that of a LoRA expert), while the larger model has 17.7M parameters (the total number of all models after 10 tasks will be comparable to that of the single large policy used in the main experiment)...

    work page 2026

  6. [7]

    Foundation models in robotics: Applications, challenges, and the future.arXiv preprint arXiv:2312.07843,

    Roya Firoozi, Johnathan Tucker, Stephen Tian, Anirudha Majumdar, Jiankai Sun, Weiyu Liu, et al. Foundation models in robotics: Applications, challenges, and the future.arXiv preprint arXiv:2312.07843,

    arXiv

  7. [9]

    Toward general- purpose robots via foundation models: A survey and meta-analysis.arXiv preprint arXiv:2312.08782,

    Yafei Hu, Quanting Xie, Vidhi Jain, Jonathan Francis, Jay Patrikar, Nikhil Keetha, et al. Toward general- purpose robots via foundation models: A survey and meta-analysis.arXiv preprint arXiv:2312.08782,

    arXiv

  8. [10]

    MoE-Loco: Mixture of experts for multitask locomotion.arXiv preprint arXiv:2503.08564,

    Runhan Huang, Shaoting Zhu, Yilun Du, and Hang Zhao. MoE-Loco: Mixture of experts for multitask locomotion.arXiv preprint arXiv:2503.08564,

    arXiv

Show all 34 references
  1. [11]

    The power of scale for parameter-efficient prompt tuning

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pp. 3045–3059,

    2021

  2. [13]

    Libero: Benchmarking knowledge transfer for lifelong robot learning.Advances in Neural Information Processing Systems, 36, 2024a

    Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning.Advances in Neural Information Processing Systems, 36, 2024a. 14 Published in Transactions on Machine Learning Research (05/2...

    2026

  3. [14]

    Progressive prompts: Continual learning for language models

    15 Published in Transactions on Machine Learning Research (05/2026) Anastasia Razdaibiedina, Yuning Mao, Rui Hou, Madian Khabsa, Mike Lewis, and Amjad Almahairi. Progressive prompts: Continual learning for language models. InThe Eleventh International Conference on Learning Re...

    2026

  4. [15]

    Don’t forget, there is more than forgetting: new metrics for continual learning.arXiv preprint arXiv:1810.13166,

    Natalia Díaz Rodríguez, Vincenzo Lomonaco, David Filliat, and Davide Maltoni. Don’t forget, there is more than forgetting: new metrics for continual learning.arXiv preprint arXiv:1810.13166,

    arXiv

  5. [17]

    Orthogonal subspace learning for language model continual learning

    Xiao Wang, Tianze Chen, Qiming Ge, Han Xia, Rong Bao, Rui Zheng, Qi Zhang, Tao Gui, and Xuanjing Huang. Orthogonal subspace learning for language model continual learning. InThe 2023 Conference on Empirical Methods in Natural Language Processing, 2023a. Yaqing Wang, Sahaj Agar...

    2023

  6. [18]

    More: Unlocking scalability in reinforcement learning for quadruped vision-language-action models

    Han Zhao, Wenxuan Song, Donglin Wang, Xinyang Tong, Pengxiang Ding, Xuelian Cheng, and Zongyuan Ge. More: Unlocking scalability in reinforcement learning for quadruped vision-language-action models. arXiv preprint arXiv:2503.08007,

    arXiv

  7. [19]

    Continual learning with pre-trained models: A survey.arXiv preprint arXiv:2401.16386,

    Dawei Zhou, Hailong Sun, et al. Continual learning with pre-trained models: A survey.arXiv preprint arXiv:2401.16386,

    arXiv

  8. [20]

    robosuite: A modular simulation framework and benchmark for robot learning.arXiv preprint arXiv:2009.12293,

    Yuke Zhu, Josiah Wong, Ajay Mandlekar, Roberto Martín-Martín, Abhishek Joshi, Soroush Nasiriany, and Yifeng Zhu. robosuite: A modular simulation framework and benchmark for robot learning.arXiv preprint arXiv:2009.12293,

    2009 arXiv

  9. [21]

    We utilize the open-source CLIP ViT/B-16 model (Radford et al.,

    17 Published in Transactions on Machine Learning Research (05/2026) A Experimental Setup A.1 Policy and Algorithm Details Policy Architecture.We use a policy consisting of vision, text, and state encoders, an input modality fusion module, a temporal transformer, and an action ...

    2026

  10. [22]

    Detailed hyperparameters of each sub-module are summarized in Table 1, most of them are naturally inherited from the open-source LIBERO benchmark and CLIP model

    as vision and text encoders, while all other components of the policy is learned from robot demonstrations. Detailed hyperparameters of each sub-module are summarized in Table 1, most of them are naturally inherited from the open-source LIBERO benchmark and CLIP model. Baselin...

    2019

  11. [24]

    establishes multifaceted prototypes through K-means clustering in the state space, where proper LoRA-based skills is retrieved upon current input state. We define the task-specific skill the same as in TAIL (Liu et al., 2024b), i.e., LoRA matrices for transformer, the fusion m...

    2026

  12. [26]

    and LoRAHub (Huang et al., 2024), which demonstrate that adjusting only the magnitude of LoRA can yield promising transfer results. The interval [0, 2] serves as a hyperparameter based on intuition to limit the magnitude, with 1 as the midpoint, representing the direct use of ...

    2024

  13. [30]

    Following the empirical analysis in Section 5.2 and illustrated in Figure 6b, we prune underactivated low-rank experts every 10 tasks to control the linear growth of the library. At the end of the 30-th task, as shown in Table 7, the number of LoRA experts in each module (the ...

    2026

  14. [31]

    Figure 13: Visualization of the action trajectories from demonstrations of LIBERO-Object benchmark

    Method Number of LoRA Experts and Parameters in Each Module Trainable Parameters Additional Storage for Replay DMPEL [15,18,17,21,17,10], 10.7M (seed=100) [18,19,17,19,17,7], 11.1M (seed=200) [20,18,20,19,18,9], 11.7M (seed=300) 1.2M 0.06GB ER / 174.1M 3.7GB 25 Published in Tr...

  15. [32]

    These tasks share the same observation and action space, but feature novel backgrounds and objects. The results in Table 8 demonstrate that DMPEL can be effectively applied to lifelong adaptation in the Robomimic benchmark, achieving better performance than IsCiL and a similar...

    2026

  16. [33]

    Figure 20: Visualization of learning curves on the LIBERO-Object benchmark

    Method FWT BWT AUC DMPEL 0.54±0.05 0.10±0.02 0.54±0.02 IsCiL 0.46±0.05 0.14±0.06 0.48±0.04 SeqFT (LoRA) 0.42±0.01 0.49±0.02 0.26±0.05 TAIL (w/ task ID) 0.42±0.01 0 0.54±0.05 29 Published in Transactions on Machine Learning Research (05/2026) Figure 19: Visualization of learnin...

    2026

  17. [100]

    When scaling to larger models, the embedding dimension increases, but modern LLMs/VLMs usually use 4k–8k embedding dimensions for 10B–100B parameter models. When scaling to ten-times-longer task sequences where κ> 100, we expect the dimension of the coefficient vector to incre...

    2026

  18. [1995]

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.π0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164,

    arXiv

  19. [2017]

    Transformer adapters for robot learning

    Anthony Liang, Ishika Singh, Karl Pertsch, and Jesse Thomason. Transformer adapters for robot learning. InCoRL 2022 Workshop on Pre-training Robot Learning,

    2022

  20. [2018]

    Progressive neural networks.arXiv preprint arXiv:1606.04671,

    Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks.arXiv preprint arXiv:1606.04671,

    arXiv

  21. [2020]

    Continual learning with tiny episodic memories.arXiv preprint arXiv:1902.10486,

    Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, et al. Continual learning with tiny episodic memories.arXiv preprint arXiv:1902.10486,

    1902 arXiv

  22. [2021]

    RT-1: Robotics transformer for real-world control at scale.arXiv preprint arXiv:2212.06817,

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, et al. RT-1: Robotics transformer for real-world control at scale.arXiv preprint arXiv:2212.06817,

    arXiv

  23. [2022]

    RT-2: Vision-language- action models transfer web knowledge to robotic control.arXiv preprint arXiv:2307.15818,

    Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, et al. RT-2: Vision-language- action models transfer web knowledge to robotic control.arXiv preprint arXiv:2307.15818,

    arXiv

  24. [2023]

    Language models are few-shot learners.Advances in Neural Information Processing Systems, 33:1877–1901,

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, et al. Language models are few-shot learners.Advances in Neural Information Processing Systems, 33:1877–1901,

    1901

  25. [2024]

    On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258,

    Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, et al. On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258,

    arXiv

  26. [2025]

    Lightweight learner for shared knowledge lifelong learning.Transactions on Machine Learning Research,

    13 Published in Transactions on Machine Learning Research (05/2026) Yunhao Ge, Yuecheng Li, Di Wu, Ao Xu, Adam M Jones, Amanda Sofie Rios, Iordanis Fostiropoulos, Po-Hsuan Huang, Zachary William Murdock, Gozde Sahin, et al. Lightweight learner for shared knowledge lifelong lea...

    2026

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.