New AI method cuts UAV relay planning time by 65%
AI transfers robot arm skills to drones, cutting urban relay setup time by 65%
Researchers from Korea University have developed Arm2Air, a cross-embodiment transfer method that converts obstacle-avoidance skills from robot arms into efficient UAV relay network configurations. The approach uses a pretrained Neural MP model to extract ordered skeletons, which are then adapted to UAV domains using Low-Rank Adaptation (LoRA).
In testing across nine complex 3D urban environments, Arm2Air reduced median end-to-end planning runtime by 64.9% compared to conventional planners. On a separate set of 30 dense urban maps, it increased bottleneck capacity by 32.6%, reduced capacity variance by 74.7%, and cut maximum hop distance by 13.2%. Notably, with just three training maps, Arm2Air achieved a 53.6% lower relay-position error than training from scratch while updating only 0.134M parameters versus 1.383M in full fine-tuning.
- Arm2Air transfers obstacle-avoidance priors from robot arms to UAVs via cross-embodiment skeleton transfer
- 64.9% faster planning and 32.6% higher bottleneck capacity in urban UAV relay networks
- Achieves better results with 10x fewer training maps (3 vs 30) and 90% fewer parameters (0.134M vs 1.383M)
Why It Matters
Enables rapid deployment of resilient drone communication networks in disaster zones or congested urban areas