acceptodds
Under review as a conference paper at ICLR 2027

Learning New Tasks via Reusable Skills: Skill-Compositional Experts for Embodied Continual Learning

Abstract

Embodied Continual Learning (ECL) aims to enable robots to continually acquire new manipulation tasks while retaining previously learned behaviors under closed-loop control. In ECL, feature drift can propagate through sequential decision-making under closed-loop control, turning representation changes into compounding behavioral deviations on previously learned tasks. A key challenge in ECL lies in structured skill reuse across continually evolving tasks, since existing methods primarily focus on skill learning without explicitly organizing them for coherent task execution. To address this issue, we propose SCE, a Skill-Compositional Experts framework for ECL. SCE builds a skill base via Compositional Skill Grounding (CSG), which decomposes task demonstrations into reusable skills. Based on this, Dual Execution-and-Transition Experts (DETE) enable new task learning through skill composition, where one branch ensures skill execution and the other supports transitions between skills for coherent behavior. Experiments on LIBERO benchmarks and real-world manipulation tasks show that SCE improves retention and overall task performance. Further feature drift analyses and ablation studies verify the effectiveness of our method. The code and real-world datasets will be released upon acceptance of the paper.

Primary area:
applications to robotics, autonomy, planning
Keywords:
embodied continual learning, skill composition, vision-language-action models, robotic manipulation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_o5FdOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SCE addresses continual robot learning through reusable skills. CSG segments demonstrations using robot-state changes and assigns semantic skill labels with a vision-language model. DETE augments the π₀ action decoder with skill-routed execution experts and a shared expert branch, combined through adaptive fusion. Training uses current-task demonstrations and replay from previous tasks. Evaluation covers four LIBERO suites and physical manipulation tasks.

Critical strengths
  1. Controlled comparisons show substantial gains. With the same π₀ backbone and replay protocol, final success over Task-MoE improves from 66.0%/61.1% to 83.9%/73.9% on LIBERO-Goal/Long, with reduced forgetting. Physical-robot final success improves from 58.9% to 80.0%.
  2. Complementary evaluations support practical utility. Component ablations, annotation-quality analysis, delayed skill arrival, and replay comparisons under a fixed update budget strengthen the empirical assessment.
Critical weaknesses
  1. The novelty of the transition branch needs clearer justification. Structurally, DETE combines shared adapters with a skill-selected adapter through adaptive fusion. The current experiments do not clearly establish what makes the shared branch a transition-specific mechanism. Removing this branch also reduces model capacity. A capacity-matched comparison with a standard shared-adapter branch would help establish the contribution beyond this architectural combination.
  2. They cover only 2–6 skills and three closely related real-robot tasks. Every new task still requires demonstrations, and the paper does not show whether reusing skills reduces the data needed to learn new tasks. Performance and computational costs with more skills also remain unclear.
Questions for authors
  1. With capacity and fusion held fixed, how does performance change without skill-probability conditioning or with non-semantic learned routing? Please specify expert configurations, routing, and parameter counts for no-CSG and Task-MoE.
  2. How do coarser or finer skill definitions affect performance and cost? Can learning curves under different demonstration budgets establish reuse benefits for new combinations of existing skills?
Reviewer_Vp6VOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates catastrophic forgetting in Vision-Language-Action models when learning new tasks under Embodied Continual Learning. The authors propose Skill-Compositional Experts, a framework that organizes continual learning around reusable skills. The method consists of two main components: Compositional Skill Grounding, which segments demonstration trajectories based on changes in robot states and constructs a semantic skill base using a VLM; and Dual Execution-and-Transition Experts, which separately model skill-specific execution and cross-skill transitions through two expert branches with adaptive fusion. The method is evaluated on four LIBERO benchmarks and three real-world robotic manipulation tasks, demonstrating improvements in continual learning performance and knowledge retention. Ablation studies, feature drift analyses, and expert routing analyses further support its effectiveness.

Critical strengths
  1. Significant improvements in continual learning performance, supported by comparisons using the same backbone. On LIBERO-Goal, SCE improves AUC from 72.3% to 82.9% and final success rate from 66.0% to 83.9% compared with Task-MoE. Notable improvements are also observed on LIBERO-Long and real-world robotic tasks. Comparisons using the same π0 backbone strengthen the experimental evidence and suggest that skill-level organization can effectively improve knowledge retention and adaptation to new tasks in continual learning.

  2. Comprehensive experimental evaluation with targeted mechanism analyses. The paper evaluates the proposed method on four LIBERO benchmarks and real-world robotic tasks, including module ablations, different skill arrival orders, replay buffer sizes, automatic skill annotation quality, and feature drift analyses. These experiments provide evidence for the effectiveness of the proposed method from multiple perspectives. In particular, the comparable performance obtained using automatically generated and manually verified skill annotations, together with the observed changes in expert fusion near skill transitions, provides valuable supporting evidence.

Critical weaknesses
  1. The novelty of the core method and the independent contribution of the skill transition mechanism are not sufficiently demonstrated. CSG primarily combines existing state-change-based trajectory segmentation with VLM-based semantic annotation, while DETE employs skill-level Top-1 expert routing, shared MoE, and adaptive fusion. In particular, the Transition Expert is routed based on the current conditional features and predicted skill distribution, without explicitly modeling the relationships between consecutive skills or introducing dedicated transition supervision. Consequently, the proposed mechanism remains relatively close to a conventional skill-conditioned MoE. Although the ablation studies demonstrate the effectiveness of the Transition Expert, they do not fully rule out the possibility that the improvements arise from additional expert capacity or conditional routing. The absence of key comparisons, such as a parameter-matched standard MoE and ablations that remove skill transition information, makes it difficult to fully establish the novelty of the central technical contribution.
Questions for authors
  1. Independent contribution of the Transition Expert.

Could the authors provide a comparison with a standard MoE matched to DETE in terms of parameter count and trainable capacity, and evaluate performance when skill transition information is removed or perturbed? Since the current Transition Expert does not employ an explicit transition-specific objective, it would be helpful to clarify how its mechanism differs from conventional skill-conditioned expert fusion.

  1. Generalization to unseen skill compositions.

Could the authors evaluate the model on previously unseen skill sequences by holding out certain skill compositions or transition relationships during training while keeping the overall skill set unchanged? Such an experiment would help assess learning efficiency and success rates when encountering novel skill combinations.

  1. Feature drift and expert mechanisms.

The paper mainly analyzes closed-loop forgetting by comparing feature drift between Sequential and SCE. Could the authors additionally provide feature drift comparisons against methods using the same backbone, such as Task-MoE, and investigate the relationship between reduced feature drift and improved task success rates?

Furthermore, in the ablation study, removing the Transition Expert results in an NBT of -6.3, whereas the full SCE achieves an NBT of 4.3. How do the authors interpret this observation in terms of the stability-plasticity trade-off?

Reviewer_DAZ2OpenReview ↗
Rating
2
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SCE for embodied continual learning. Compositional skill grounding divides demonstrations into temporal segments and assigns reusable skill labels with a VLM. The action decoder combines one selected execution expert with a weighted mixture of Transition experts. An adaptive gate combines their outputs. The method keeps the original action decoder fixed and trains adapters, routing modules, and vision-language LoRA parameters. Experiments cover four LIBERO suites and three sequential real-world tasks.

Critical strengths

The reviewer thinks that the authors have evaluated multiple practical concerns, including larger replay buffers, delayed skill arrival, and automatic versus verified annotations. These experiments provide useful evidence beyond the two main suites. The larger-replay results also identify a resource trade-off rather than presenting SCE as uniformly superior.

Critical weaknesses

The reviewer does not think that the experiments establish the specific execution-and-transition mechanism that motivates the method. Specifically, Equations (3)–(6) seem to describe a hard-routed skill adapter and a soft mixture of shared adapters. However, the method does not explicitly represent the previous skill, a skill pair, or a temporal change in the skill distribution. The resulting performance drop combines changes in model capacity, routing, and the proposed transition function. The reviewer thinks that the changes in fusion weights near skill boundaries support the interpretation, but they do not establish its causal role.

Questions for authors
  1. What exactly changes in the “without CSG” ablation? The reviewer thinks that this variant remove only the skill-classification loss, but not sure.
  2. Does the feature-drift analysis isolate the effect of the proposed structure?
Figure 4 compares SCE with Sequential, which also differs in replay. The paper should report the same measurements for ER and Task-MoE.
Which modules can each comparison method update? Can the authors report absolute feature distances and action differences? Smaller feature drift and smaller rollout divergence are consistent with retention, but they do not establish which component causes the improvement.
  3. How does skill-space expansion preserve existing routing behavior? It seems appending classifier outputs changes the softmax normalization, then adding transition-projection columns can also change the shared router’s output. Namely, preserving existing parameters does not necessarily preserve the policy on old inputs.
  4. How reliable are the real-world and transition-specific results? The real-world study uses three closely related tasks and 30 trials per task at each stage.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.