VLA-Forget: Vision-Language-Action Unlearning for Embodied Foundation Models

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

VLA-Forget is a hybrid unlearning framework designed for Vision-Language-Action (VLA) models, which serve as embodied foundation models for robotic manipulation. It addresses the challenge of removing unsafe, spurious, or privacy-sensitive behaviors without degrading core functionalities like perception, language grounding, and action control. Unlike conventional methods, VLA-Forget tackles knowledge distributed across perception, alignment, and reasoning/action layers, where partial unlearning is often insufficient. The framework 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. Evaluations show 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 unlearning baselines.

Key takeaway

For Robotics Engineers deploying Vision-Language-Action models, ensuring the removal of unsafe or privacy-sensitive behaviors is critical. VLA-Forget offers a robust, hybrid unlearning framework that effectively targets undesirable knowledge across model layers while preserving essential perception and reasoning capabilities. You should consider its multi-objective optimization approach to achieve efficient and utility-preserving unlearning in your embodied foundation models.

Key insights

Unlearning in embodied Vision-Language-Action models requires a multi-layered, hybrid approach to preserve utility.

Principles

• Undesirable knowledge distributes across VLA model layers.
• Partial unlearning is insufficient for embodied VLA models.
• Conventional unlearning baselines degrade utility in VLA settings.

Method

VLA-Forget jointly optimizes targeted forgetting, perceptual preservation, and reasoning retention through staged updates across the visual encoder, projector, and upper transformer blocks.

In practice

• Remove unsafe behaviors from robotic policies.
• Address privacy concerns in embodied AI models.

Topics

• Vision-Language-Action Models
• Embodied AI
• Machine Unlearning
• Robotic Manipulation
• Foundation Models
• VLA-Forget

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• $μ$VLA: On Recurrent Memory for Partially Observable Manipulation in VLA Models — Isolating minimal recurrence in VLA models significantly enhances performance in partially observable robotic manipulation.
• Motion-Focused Latent Action Enables Cross-Embodiment VLA Training from Human EgoVideos — Unlabeled human videos can effectively train VLA models by disentangling motion and action intent.
• Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think — Vision-Language-Action models can be significantly compressed by removing redundant layers identified via a training-free method.
• Retrieve, Don't Retrain: Extending Vision Language Action Models to New Tasks at Test Time — Adapting Vision-Language-Action models to new tasks can be achieved via demonstration retrieval, not costly per-task fine-tuning.
• VisualThink-VLA: Visual Intermediate Reasoning for Effective and Low-Latency Vision-Language-Action Policies — Visual intermediate reasoning, not textual chain-of-thought, enables effective and low-latency vision-language-action policies.
• LARA: Latent Action Representation Alignment for Vision-Language-Action Models — LARA jointly optimizes Latent Action Models and Vision-Language-Action models via representation alignment for improved robot control.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.