VLA-Forget: Vision-Language-Action Unlearning for Embodied Foundation Models (April 2026)

Free

First unlearning method for VLA models — removes target behaviors while preserving general capabilities; introduces forget/retain/boundary splits and real-robot OXE benchmarks

FreeFree tier

Type

Open Source

Visit Website GitHub (40 stars)

About VLA-Forget: Vision-Language-Action Unlearning for Embodied Foundation Models (April 2026)

VLA-Forget is a hybrid unlearning framework designed for Vision-Language-Action (VLA) models used in robotic manipulation. It addresses the challenge of removing unsafe, spurious, or privacy-sensitive behaviors from VLA policies without degrading perception, language grounding, or action control. The method combines ratio-aware selective editing for perception and cross-modal specificity with layer-selective reasoning/action unlearning. It jointly optimizes three objectives—targeted forgetting, perceptual preservation, and reasoning retention—through staged updates over the visual encoder, projector, and upper action-generating transformer blocks. Evaluated on forget-set behavior probes and retain-task evaluations, VLA-Forget improves forgetting efficacy by 10%, preserves perceptual specificity by 22%, retains reasoning and task success by 9%, and reduces post-quantization recovery by 55% relative to strong baselines. The framework also introduces novel forget/retain/boundary data splits and real-robot OXE benchmarks for systematic evaluation.

Key Features

Hybrid unlearning framework combining ratio-aware selective editing and layer-selective unlearning

Joint optimization of targeted forgetting, perceptual preservation, and reasoning retention

Staged updates over visual encoder, cross-modal projector, and upper transformer blocks

Introduces forget/retain/boundary data splits for systematic evaluation

Real-robot OXE benchmarks for assessing unlearning in embodied settings

Pros & Cons

Pros

Achieves 10% improvement in forgetting efficacy over strong baselines
Preserves perceptual specificity by 22% while forgetting target behaviors
Retains reasoning and task success by 9%
Reduces post-quantization recovery of forgotten behaviors by 55%

Cons

Designed specifically for OpenVLA-style policies and may require adaptation for other VLA architectures
Requires access to forget, retain, and boundary data splits which may not be readily available

Best For

Removing unsafe or undesirable behaviors from pretrained VLA robotic manipulation modelsEliminating spurious correlations or privacy-sensitive memorized actions from embodied foundation modelsSelective unlearning in VLA policies to meet safety or regulatory requirements without full retraining

FAQ

What is VLA-Forget?

VLA-Forget is a hybrid unlearning framework for vision-language-action (VLA) models that removes unsafe, spurious, or privacy-sensitive behaviors while preserving perception, language grounding, and action control capabilities.

How does VLA-Forget work?

It combines ratio-aware selective editing for perception and cross-modal specificity with layer-selective reasoning/action unlearning. The method jointly optimizes three objectives—targeted forgetting, perceptual preservation, and reasoning retention—through staged updates over the visual encoder, projector, and upper action-generating transformer blocks.

What improvements does VLA-Forget demonstrate?

Relative to strong baselines, VLA-Forget improves forgetting efficacy by 10%, preserves perceptual specificity by 22%, retains reasoning and task success by 9%, and reduces post-quantization recovery of forgotten behaviors by 55%.

Is VLA-Forget model-agnostic?

The framework is designed for OpenVLA-style policies but the principles of hybrid unlearning could be adapted to other VLA architectures.