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

Ralf Römer^1,*, Yi Zhang^1,*, Angela P. Schoellig¹,

¹Technical University of Munich

The official code repository for "CLARE: Continual Learning for Vision-Language-Action Models via Autonomous Adapter Routing and Expansion".

Abstract: To teach robots complex manipulation tasks, it is now a common practice 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 non-exemplar continual learning with VLAs. CLARE introduces lightweight modular adapters into selected feedforward layers 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, we show that CLARE achieves high performance on new tasks without catastrophic forgetting of earlier tasks, significantly outperforming even exemplar-based methods.

Project Structure

This codebase is built on top of two open-source frameworks from Hugging Face:

• lerobot_lsy/: Modified version of LeRobot for designing, training, and fine-tuning Vision-Language-Action (VLA) models
• peft_lsy/: Modified version of PEFT implementing the CLARE algorithm as a LoRA-compatible adapter

Pre-trained Checkpoints & Datasets

We provide pre-trained model checkpoints and LIBERO datasets on 🤗 Hugging Face: huggingface.co/continuallearning

Available resources:

• dit_flow_mt_libero_90_pretrain and dit_mt_libero_90_pretrain: DiT-Dec and DiT-EncDec base checkpoints pretrained on LIBERO-90 as described in the paper
• libero_10_image_task_0 to libero_10_image_task_9: LIBERO-10 benchmark datasets

Installation

Prerequisites

• Python 3.8+
• CUDA-compatible GPU (recommended)
• Conda or Miniconda

Setup

1. Create and activate a conda environment

   conda create -n clare python=3.10
   conda activate clare

2. Install PEFT-LSY in editable mode

   cd peft_lsy
   pip install -e .
   cd ..

3. Install LeRobot-LSY in editable mode

   cd lerobot_lsy
   pip install -e .
   cd ..

4. Install additional dependencies (if needed)

   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

PEFT Configuration

CLARE uses a custom PEFT adapter configuration. Below is an example configuration for the CLARE adapter for the DiT-Dec policy:

{
  "peft_type": "CLARE",
  "task_type": null,
  "auto_mapping": {
    "base_model_class": "PeftWrapperPolicy",
    "parent_library": "__main__"
  },
  "base_model_name_or_path": null,
  "revision": null,
  "target_modules": ".*velocity_net.cond_proj",
  "inference_mode": true,
  "batch_first": true,
  "num_learned_task": 0,
  "feature_dim": 2576,
  "out_feature_dim": 512,
  "use_trainable_copy": false,
  "add_zero_init_conv_layer": false,
  "structure": {},
  "discriminator_cfg": {
    "type": "autoencoder",
    "batch_first": true,
    "feature_dim": 2576,
    "feature_fusion": false,
    "fused_feature_dim": null,
    "hidden_dim": 256,
    "latent_dim": 128,
    "num_tokens": 16,
    "lora_rank": 32,
    "lora_alpha": 32,
    "use_lora": false,
    "use_momentum": true,
    "momentum": 0.1,
    "max_batches_tracked": 2000
  },
  "func_adapter_cfg": {
    "hidden_dim": 1024,
    "lora_rank": 32,
    "lora_alpha": 32,
    "use_lora": false
  }
}

Key Configuration Parameters

• target_modules: Regular expression pattern matching the model layers where adapters will be applied
• feature_dim: Dimension of input features for the adapter
• out_feature_dim: Output dimension after adapter transformation
• discriminator_cfg: Configuration for the autoencoder-based routing mechanism
  • type: Type of discriminator (autoencoder or other)
  • hidden_dim: Hidden layer dimension
  • latent_dim: Latent space dimension for the autoencoder
  • use_momentum: Whether to use momentum for feature statistics
• func_adapter_cfg: Configuration for functional adapter modules
• num_learned_task: Number of tasks learned so far

Usage

Training with CLARE on LIBERO Benchmark

The main training script is located at lerobot_lsy/src/lerobot/scripts/clare.py. Below is an example command for the first stage of continual learning on the LIBERO-10 benchmark:

python ./lerobot_lsy/src/lerobot/scripts/clare.py \
    --seed=42 \
    --job_name=clare_libero_10_task_0 \
    --output_dir=./outputs/clare_libero_10_task_0 \
    --dataset.repo_id=continuallearning/libero_10_image_task_0 \
    --policy.path=continuallearning/dit_mt_libero_90_pretrain \
    --policy.push_to_hub=false \
    --batch_size=32 \
    --num_workers=16 \
    --steps=20000 \
    --env.type=libero \
    --env.task=Libero_10_Task_0 \
    --eval.batch_size=20 \
    --eval.n_episodes=100 \
    --eval.max_episodes_rendered=100 \
    --eval_freq=200000 \
    --save_freq=20000 \
    --log_freq=100 \
    --peft_cfg_path=./peft_lsy/config \
    --expand_threshold=10.00 \
    --detect_distribution_shift_steps=200 \
    --detect_distribution_shift_batch_size=32 \
    --detect_distribution_shift_num_workers=16 \
    --detect_distribution_shift_log_freq=10 \
    --train_discriminators_steps=2000 \
    --train_discriminators_batch_size=32 \
    --train_discriminators_num_workers=16 \
    --train_discriminators_log_freq=50 \
    --train_discriminators_eval_freq=2000 \
    --train_discriminators_save_freq=2000 \
    --wandb.enable=true \
    --wandb.disable_artifact=true \
    --wandb.project=clare_experiments \
    --wandb.entity=<your-wandb-entity>

You can also run a full continual learning experiment for DiT-Dec using the provided bash script:

bash bash/dit_dec.sh

Key Training Arguments

Dataset & Model

• --dataset.repo_id: Hugging Face dataset repository ID
• --dataset.root: Local path to the dataset
• --policy.path: Path to the pre-trained VLA model

Training Configuration

• --batch_size: Training batch size
• --num_workers: Number of data loading workers
• --steps: Total training steps
• --seed: Random seed for reproducibility

CLARE-Specific Parameters

• --peft_cfg_path: Path to the PEFT configuration JSON file
• --expand_threshold: Threshold for autonomous adapter expansion (based on feature similarity)
• --detect_distribution_shift_steps: Steps for distribution shift detection
• --train_discriminators_steps: Training steps for the autoencoder-based routing mechanism

Evaluation

• --env.type: Environment type (e.g., libero, aloha)
• --env.task: Specific task name
• --eval.n_episodes: Number of evaluation episodes
• --eval_freq: Frequency of evaluation (in training steps)

Logging

• --wandb.enable: Enable Weights & Biases logging
• --wandb.project: W&B project name
• --log_freq: Logging frequency

Evaluation

You can run evaluations in LIBERO by either directly using the latest LeRobot version, which includes LIBERO, or by installing LIBERO separately:

cd ..
git clone git@github.com:ZhangYi1999/gym-libero.git
cd gym-libero
pip install -e .
python test/test_gym_libero.py  # Validate installation

Citation

If you find this work useful, please consider citing our paper:

@article{clare,
  title={CLARE: Continual Learning for Vision-Language-Action Models via Autonomous Adapter Routing and Expansion},
  author={Ralf R{\"o}mer and Yi Zhang and Angela P. Schoellig},
  journal={arXiv preprint arXiv:2601.09512},
  year={2026}
}

License

This project is released under the same license as the original LeRobot and PEFT frameworks.

Acknowledgments

This work builds upon:

• LeRobot by Hugging Face
• PEFT by Hugging Face
• LIBERO by Bo Liu et al.

We thank the authors of these projects for their open-source contributions.