PaperlayerView exact revision on arXiv

Source-linked AI summary

CLARE: Continual Learning for Vision-Language-Action Models via Autonomous Adapter Routing and Expansion

Ralf Römer, Yi Zhang, Yuming Li, Angela P. Schoellig

arXiv:2601.09512v2cs.ROcs.LG

TL;DR

VLAs must adapt to sequentially arriving tasks while retaining prior skills, but full fine-tuning can cause forgetting and existing methods may require exemplars or task identifiers. CLARE adds lightweight adapters with autonomous routing and selective expansion, achieving strong continual-learning performance across simulated and hardware tasks, including 63.3% AUC and −2.9% NBT in hardware experiments.

  • Problem

    Robots need exemplar-free continual learning because sequential adaptation must preserve prior knowledge when old data may be unavailable and task identifiers may not be accessible.

  • Method

    CLARE injects lightweight adapters into selected VLA modules, expands them when feature statistics indicate novelty, and routes adapters autonomously using feature similarity.

  • Results

    Across LIBERO suites and five hardware tasks, CLARE maintains high success without forgetting and outperforms strong baselines; hardware performance reaches 63.3% AUC and −2.9% NBT.

  • Takeaways & Limitations

    CLARE supports long-term VLA deployment by adding capacity only when needed while retaining prior representations without stored exemplars or task IDs.

Abstract

from arXiv · show

To teach robots complex manipulation tasks, a common approach is to fine-tune a pre-trained vision-language-action model (VLA) on task-specific data. However, since this recipe updates existing representations, it is unsuitable for long-term operation in the real world, where robots must continually adapt to new tasks and environments while retaining the knowledge they have already acquired. Existing continual learning methods for robotics commonly require storing previous data (exemplars), struggle with long task sequences, or rely on task identifiers for deployment. To address these limitations, we propose CLARE, a general, parameter-efficient framework for exemplar-free continual learning with VLAs. CLARE introduces lightweight modular adapters into selected VLA modules and autonomously expands the model only where necessary when learning a new task, guided by layer-wise feature similarity. During deployment, an autoencoder-based routing mechanism dynamically activates the most relevant adapters without requiring task labels. Through extensive experiments on the LIBERO benchmark and five real-world tasks, we show that CLARE achieves high performance on new tasks without catastrophic forgetting of earlier tasks, significantly outperforming even exemplar-based methods. Code, data, and videos are available at our website: https://tum-lsy.github.io/clare.

I. INTRODUCTION

CLARE targets exemplar-free continual learning for VLAs, enabling robots to acquire new skills while preserving prior knowledge during long-term operation. It combines modular adapters, autonomous routing, and dynamic expansion to avoid task identifiers and limit parameter growth.

  • Motivation: Robots must continually learn new skills in changing environments without sacrificing previously acquired capabilities.This remains an open challenge for long-term operation in homes, hospitals, and warehouses.
  • Limitations of Existing Approaches: Iterative full-model fine-tuning updates shared representations and can cause catastrophic forgetting of semantic grounding and old-task policy performance.This makes current VLAs ill-suited to non-stationary deployment.
  • Limitations of Existing Approaches: Experience replay mitigates forgetting but requires storing and accessing past data, creating memory, computation, privacy, and sample-selection burdens.These constraints motivate exemplar-free continual learning for robotics.
  • Limitations of Existing Approaches: Modular and expandable methods add capacity instead of overwriting shared representations, but often require unavailable task identifiers and assume tasks are known in advance.Their robotics applications have largely focused on multitask learning with all tasks and data available beforehand.
  • CLARE: CLARE injects lightweight adapters into selected VLA modules and expands only when feature statistics indicate substantial novelty.It aims to preserve pretrained representations while adding task-specific capacity as needed.
  • CLARE: CLARE routes inference autonomously using feature similarity, increases parameters by about 2% per task, and outperforms continual-learning baselines in simulation and real-world settings.The design is reported as architecture-agnostic and applicable to larger VLAs.

2) Continual Learning:

The continual-learning setting presents sequentially arriving tasks with unknown total length, requiring adaptation from only the current demonstrations and previous model parameters. CLARE is framed against replay, regularization, and architectural methods that trade memory, capacity, or autonomy in different ways.

  • Continual Learning: Continual learning requires acquiring new skills while retaining previously learned capabilities and preserving plasticity.This is described as a hard problem in deep learning and robotics.
  • Continual Learning: Experience replay mixes stored past examples with new data, but storing exemplars may be infeasible because of data scarcity and safety concerns.Replay also introduces storage and access requirements in lifelong robotic learning.
  • Prior Methods: Regularization methods constrain updates to important weights, while PackNet reallocates pruned parameters; both face restrictions from a fixed initial parameter set over long task sequences.Architectural methods address capacity bottlenecks by injecting new parameters or modules.
  • Robotics Setting: Some robotics methods use task identifiers or schedulers for deployment, preventing fully autonomous operation when task identity is unavailable.SDP is cited as requiring oracle task identifiers.
  • Problem Formulation: The target setting supplies sequential task demonstrations while earlier datasets are unavailable, so adaptation must use the previous model and current task data.The total number of tasks is unknown, and the policy should retain pretrained knowledge and earlier task skills.

IV. METHODOLOGY

CLARE is presented as a continual-learning framework for VLAs, with its training and inference strategies summarized in two algorithms.

  • IV. METHODOLOGY: CLARE’s methodology section introduces the proposed continual-learning framework and summarizes its training and inference strategies.Algorithms 1 and 2 provide the corresponding procedural summaries.

A. Base Policy

The base policy is trained with flow matching to generate action chunks from multimodal observations. During deployment, discriminator reconstruction errors select the relevant adapter, whose output is combined with the original module.

  • Base Policy: CLARE trains the policy using flow matching for high-dimensional, multimodal demonstration data.The supplied passage identifies flow matching as the training method.
  • Autonomous Routing: During deployment, discriminators compute reconstruction errors for an input feature and select the most relevant adapter.The selected adapter is activated through the routing procedure.
  • Autonomous Routing: The routed adapter output is summed with the output of the original module.This preserves the original module’s contribution while incorporating the selected adapter.
  • Base Policy: The policy learns a vector field that transports samples from a simple base distribution, such as Gaussian noise, to the target action-chunk distribution.Euler integration then generates action chunks from Gaussian noise conditioned on observations.

B. Modularized Adapters

CLARE adds lightweight, task-specific adapters as parallel side branches while freezing existing parameters, then routes observations to relevant adapters without task identifiers.

  • Modularized adapters: CLARE continually injects parameters into selected VLA modules while keeping the base model structure and existing representations intact.Adapters are added as parallel side branches, preserving layer inputs and outputs.
  • Modularized adapters: At each continual-learning stage, at most one lightweight adapter is added to each expandable layer, and only newly added adapters are trained.The remaining model parameters and previously learned adapters are frozen.
  • Autonomous routing: During deployment, CLARE selects an adapter autonomously for each expandable layer from an expanding adapter set, without task labels.The router uses current layer features as input and pairs each layer with expandable autoencoder discriminators.
  • Autonomous routing: The router chooses the adapter linked to the discriminator with the smallest reconstruction error for the current feature.Discriminators are trained so their reconstruction errors are lower on features from their corresponding adapter distributions.
  • Autonomous routing: CLARE trains new adapters first with flow matching, then freezes all parameters except new discriminators and trains them with reconstruction loss.This two-stage procedure addresses feature-distribution changes caused by training adapters in shallower layers.

D. Dynamic Expansion

CLARE expands only layers whose new-task features are sufficiently novel, while adding discriminators when needed to stabilize routing and limit parameter growth.

  • Dynamic expansion: If a layer’s z-scores exceed threshold γ, CLARE adds a new adapter and discriminator; otherwise, it adds only a discriminator linked to an existing adapter.The discriminator is linked to the most relevant adapter when expansion is unnecessary.
  • Dynamic expansion: CLARE expands a layer only when the new task’s features deviate substantially from all previously learned tasks.Reconstruction losses are normalized before comparing them across discriminators.
  • Dynamic expansion: CLARE’s dynamic expansion produces a memory-efficient, sublinear increase in adapter parameters rather than expanding every layer for every task.Adding adapters to all expandable layers would cause excessive linear parameter growth and limit knowledge sharing.
  • Parameter efficiency: The injected modules are much smaller than the approximately 200M-parameter base policy.This supports the parameter-efficient design used for continual learning.
  • Routing stability: A new discriminator is still attached to an unexpanded layer so later routing changes do not shift downstream features unpredictably.Without the auxiliary discriminator, revisiting an earlier task could activate a newer adapter and cause task failure.
  • Dynamic expansion: CLARE adds at least some new parameters per task and defaults to expanding all layers at the first stage.If no layer requires expansion, an adapter is still added to the shallowest expandable layer because shallow layers typically show stronger task distribution shifts.

V. EVALUATION

The evaluation examines CLARE’s layer choices, continual-learning performance, skill reuse, and computational overhead across simulation and real-world settings.

  • Research questions: The experiments ask which VLA layers are best suited for adapter expansion.This is evaluation question Q1.
  • Research questions: The evaluation measures whether CLARE learns new tasks while preserving performance on previous tasks.This is evaluation question Q2.
  • Research questions: The experiments test whether autonomous dynamic expansion can reuse relevant skills from previous tasks.This is evaluation question Q3.
  • Research questions: The evaluation also measures CLARE’s computational overhead.This is evaluation question Q4.
  • Model architecture: The study uses a base VLA architecture whose adapter insertion locations are indicated as dashed blocks in the model diagram.The figure presents the base policy architecture and identifies modules selected for CLARE adapters.

A. Experimental Setup

The experiments evaluate CLARE in simulation and hardware settings using sequential manipulation tasks, with ablations examining policy architecture and long-task scalability.

  • Simulation: Simulation uses a Panda manipulator with 50 demonstrations per task, pre-trained on LIBERO-90 and evaluated on 10 sequential tasks from three LIBERO suites.The evaluation uses the LIBERO-Long, LIBERO-Goal, and LIBERO-Spatial suites.
  • Hardware: Hardware experiments use an FR3 manipulator across five tasks involving bowls, stacking, a Moka pot, a drawer, and LEGO.The tasks cover varied object interactions and dynamics.
  • Policy: The base VLA is a six-layer decoder-only diffusion transformer with frozen DINOv2 and CLIP encoders, while adapters can be inserted into encoder projections and decoder layers.The architecture uses multimodal token projections and AdaLN conditioning.
  • Ablations and scalability: The ablation examines adding adapters to observation-encoding modules, while LIBERO-40 tests continual learning across 40 tasks.The setup therefore evaluates both module placement and long-sequence scalability.

3) Metrics:

The study evaluates continual learning with metrics for overall performance, new-task learning, and forgetting, then compares CLARE with several baselines across increasingly long task sequences.

  • Metrics: AUC measures overall performance on new and old tasks, FWT measures new-task learning ability, and NBT measures forgetting, with lower NBT being better.All numerical results are reported in percentage points.
  • Evaluation protocol: After each stage, the policy is evaluated on all previously learned tasks using 100 simulation rollouts per task and results averaged across three seeds.Real-world evaluation uses 10 rollouts per task.
  • Baselines: The baseline set includes sequential full fine-tuning, LoRA, PackNet, and other continual-learning methods without oracle task identifiers.SeqFFT fine-tunes the whole model, whereas SeqLoRA merges adapters after each task.
  • Simulation results: CLARE achieves the highest AUC, outperforming ER by about 10 to 14 percentage points while achieving comparable FWT to SeqFFT and ER.The comparison indicates strong overall retention and new-task learning.
  • Long-term scalability: CLARE achieves approximately zero NBT, and on LIBERO-40 it retains 40 tasks while ER reaches an NBT of 23%.The long-sequence result contrasts CLARE’s retention with ER’s forgetting on several tasks.

1. BOWL

Hardware experiments assess CLARE on real manipulation tasks, reporting strong continual-learning performance with low overhead and near-zero forgetting.

  • Hardware evaluation: The hardware evaluation uses an FR3 manipulator and compares CLARE with SeqFFT, SeqLoRA, and ER.The experiments use five real-world manipulation tasks.
  • Results: CLARE achieves an AUC of 63% and 12 percentage points higher than ER, with an NBT of -2.9%.The result combines overall performance with near-zero forgetting.
  • Adapter placement: CLARE adds adapters to observation projections and AdaLN scale-and-shift modules based on the importance of observation encoding.This placement follows the simulation ablation.
  • Efficiency: Inference-time overhead stays below 3 ms relative to the base policy, while GPU VRAM increases by about 2% per task.The dedicated lightweight modules are reported as more storage-efficient than storing previous ER data.
  • Conclusion: Across the hardware experiments, CLARE retains prior representations while learning new tasks under real-world operating conditions.The conclusion is supported by the reported AUC and near-zero NBT.

VII. APPENDIX

The extended experiments examine task diversity, robustness across baselines, router switching, adapter placement, and real-world performance under changing conditions.

  • Task diversity: The five real-world tasks vary in object properties, grasp configurations, interaction dynamics, and multistage strategy requirements.Examples include different weights, gripper angles, nonlinear drawer friction, and LEGO’s three-stage procedure.
  • Routing: Changing the language command from placing LEGO to closing the drawer causes the router to immediately switch to the correct adapter during execution.The robot consequently stops approaching the block and moves toward the drawer.
  • Ablation: Expanding only the encoder performs similarly to expanding both encoder and decoder modules, but substantially better than expanding only the decoder.The findings identify observation-conditioning modules as crucial for adapter injection.
  • Extended discussion: Hardware results reach an AUC of 63.3% and near-zero NBT of −2.9% despite daylight changes, reflections, camera drift, and physically distinct tasks.The expansion strategy is reported not to bottleneck physical dynamics in this setup.
  • Benchmark comparison: On LIBERO-Long, CLARE achieves a higher overall success rate and mitigates catastrophic forgetting more effectively than five baselines.The curves average three random seeds, with shaded standard deviations.
  • Scope: The hardware expansion strategy assumes pretraining on sufficiently large robot-demonstration datasets covering diverse motion patterns.The authors identify this pretraining condition as a prerequisite for the reported physical setup.
Loading 2601.09512v2…