Federated vision transformer enabling collaborative object manipulation in bandwidth-limited multi-robot systems

Multi-robot collaborative object manipulation systems have broad application prospects in intelligent manufacturing and warehouse logistics, but the limited computational power of edge devices and wireless communication bandwidth constraints severely restrict the distributed deployment of deep learning models. This paper proposes a multi-robot collaborative control framework integrating federated learning and Vision Transformer, aiming to achieve efficient collaborative manipulation under bandwidth-constrained conditions. At the model level, a lightweight Vision Transformer architecture based on depthwise separable convolution tokenization and bottleneck feed-forward networks is designed, achieving 78.4±1.2% object detection accuracy and 34.7 FPS inference speed with only 5.4M parameters. At the communication level, bandwidth-adaptive gradient compression and asynchronous aggregation strategies are constructed, enabling importance-based transmission through gradient importance scoring, reducing communication overhead by 87.3% while maintaining model accuracy loss below 2%. At the collaboration level, a distributed task allocation and trajectory coordination mechanism based on shared latent space is established, enabling multiple robots to achieve safe and efficient collaborative operations under limited communication conditions. Simulation and real robot experiments demonstrate that the proposed method achieves a 92.7±1.8% task success rate (mean ± 95% CI, n = 5 runs) in industrial parts sorting scenarios, with a three-robot collaboration speedup ratio of 2.62±0.18 and a 74.5%±4.2% reduction in collision occurrences. It should be noted that this framework achieves data locality through gradient-only communication rather than formal differential privacy guarantees. This framework provides an effective technical solution for multi-robot system deployment in bandwidth-constrained environments.