MIT's DGPPO enables safe multi-drone transport with zero-shot sim-to-real
A single policy trained in simulation handles real drone teams of varying sizes safely.
Researchers from MIT (Jaeyoun Choi, Oswin So, Songyuan Zhang, Cooper Taylor, Chuchu Fan) published a paper in IEEE Robotics and Automation Letters 2026 introducing a learning-based framework for safe and scalable multi-drone cooperative payload transport. Their approach, called Discrete Graph Control Barrier Function Proximal Policy Optimization (DGPPO), combines reinforcement learning with control barrier functions (CBFs) to ensure safety during complex coupled dynamics among drones, cables, and payloads. To make multi-drone training computationally tractable, they designed a minimal 2D abstraction that preserves task-relevant coupling while enabling large-scale learning. Through domain randomization over team size and physical parameters, the policy learns robust behaviors that transfer directly from simulation to real hardware without any fine-tuning.
Extensive real-world evaluations demonstrate that a single learned policy can generalize across varying team sizes (e.g., 2 to 4 drones carrying a payload) and handle dynamic environments where other drone teams act as moving obstacles. The zero-shot sim-to-real capability eliminates the need for expensive real-world data collection or iterative tuning, making the system practical for rapid deployment in logistics, construction, and disaster response. The work represents a significant step toward safe autonomous drone swarms that can collaborate on heavy lifting tasks in unstructured environments.
- DGPPO uses control barrier functions (CBFs) inside a graph-based PPO framework to enforce safety constraints during multi-drone payload transport.
- The policy is trained entirely in simulation with domain randomization and transfers zero-shot to real drones without any fine-tuning.
- Real-world experiments show it generalizes across team sizes (2–4 drones) and safely navigates around other moving drone teams.
Why It Matters
Enables safe, scalable autonomous drone swarms for logistics and construction without costly real-world training.