ForgeVLA: Federated Vision-Language-Action Learning without Language Annotations

ForgeVLA introduces a federated learning framework that enables Vision-Language-Action models to train on distributed robot data without centralizing sensitive information or requiring manual language annotations. The system uses embodied instruction classifiers to automatically generate missing language labels and addresses vision-language feature collapse through contrastive learning and adaptive aggregation.

ForgeVLA addresses a critical bottleneck in scaling robotic AI systems: the prohibitive cost of collecting and annotating training data while respecting data privacy constraints across distributed deployments. Traditional VLA models require expensive human annotations and centralized data aggregation, which is impractical for robots deployed across hospitals, factories, warehouses, and research institutions. This research tackles the problem by enabling federated learning where robots collaboratively improve shared models while keeping raw data local.

The framework's innovation centers on two technical contributions. First, it recovers the missing language modality by deploying lightweight embodied instruction classifiers on each client robot, automatically mapping observed vision-action pairs to predefined task instructions. This eliminates the annotation bottleneck entirely. Second, the authors identify and solve vision-language feature collapse—a phenomenon where multimodal representations lose discriminative power during distributed training. They combine client-side contrastive planning losses with server-side adaptive aggregation to maintain task-relevant feature distinctions.

This advancement has substantial implications for the robotics and embodied AI industries. By reducing annotation costs and enabling privacy-preserving collaborative learning, ForgeVLA lowers barriers to training next-generation general-purpose robotic systems. Organizations can leverage their proprietary deployment data to build competitive advantage without exposing sensitive operations. The framework's demonstrated performance improvements over baselines suggest practical viability for real-world deployments.

Looking ahead, the success of federated VLA training could accelerate adoption of foundation models in robotics, similar to how federated learning transformed mobile AI. Key areas to monitor include scaling to heterogeneous robot morphologies, extending to more complex instruction sets, and integration with commercial robotics platforms.

• →ForgeVLA enables federated VLA training without centralizing raw data or requiring manual language annotations through automated instruction classification.
• →The framework addresses vision-language feature collapse using contrastive planning and adaptive aggregation strategies.
• →Privacy-preserving federated learning lowers barriers for organizations to collaboratively improve robotic models using proprietary deployment data.
• →Performance benchmarks demonstrate significant improvements over existing baselines across multiple evaluation datasets.
• →The approach could accelerate foundation model adoption in robotics by reducing annotation costs and enabling collaborative learning.