6G is expected to provide the key infrastructure for large‑scale embodied intelligence, where heterogeneous robots collaborate via low‑latency connectivity, edge intelligence, and distributed sensing. Vision‑language‑action (VLA) models form a unified closed‑loop policy that integrates visual perception, language understanding, and action generation. However, training and adapting VLA models for distributed robotic agents raises challenges in privacy protection, communication efficiency, and model heterogeneity. Existing federated learning (FL) approaches overlook the intrinsic differences among vision, language, and action pathways in parameter scale, privacy exposure, update dynamics, and tolerance to compression or perturbation.
To address this, we propose FedMVLA, a modality‑decoupled FL framework tailored for privacy‑preserving embodied intelligence in 6G networks. FedMVLA comprises three mechanisms:
- MAFA (Modality‑Aware Federated Aggregation): applies modality‑specific aggregation weights and frequencies to ensure convergence of vision, language, and action parameters individually.
- MAPA (Modality‑Aware Privacy Allocation): allocates differential privacy budgets $\epsilon$ according to each modality’s privacy risk, e.g., a stricter $\epsilon_{action}$ for the action stream.
- MACO (Modality‑Aware Communication Compression): employs high‑compression quantization/sparsification for vision and language parameters while keeping low compression for the action stream to meet real‑time requirements.
FedMVLA also introduces a modality‑sliced transport design that routes the precision‑critical action stream through a protected ultra‑reliable low‑latency (URLLC) slice, while other modalities use the conventional enhanced mobile broadband (eMBB) slice.
A case study is conducted on a 3GPP‑based wireless substrate, accounting for fading, co‑channel interference, and malicious jamming. Results show:
- Task success rate of 84.8%, surpassing FedAvg by 22.2 percentage points.
- Success rate continues to improve when scaling to 128 clients across eight cells.
- Per‑client uplink model‑update payload is reduced by about 95.6% (≈96%), and the 95th percentile uplink completion time remains near 1.5 s.
These findings demonstrate that modality‑aware aggregation, privacy allocation, and communication compression can dramatically enhance the efficiency and reliability of distributed robotic collaboration while preserving privacy and real‑time performance.
Review