Stanford enables robotic arms with AI to directly fly drones: pick up objects and navigate autonomously without retraining

What Happened

The Stanford team did something interesting: they took a VLA model trained entirely on fixed robotic arm data and had it fly drones and grab objects. Their solution is called AirVLA, based on π₀ VLA, which added a layer of “payload-aware” physical guidance to adapt to flight dynamics, and then used 3D Gaussian Splatting to generate synthetic data to supplement navigation samples.

What Numbers Came Out

• Navigation Success Rate: 100%
• Grasp/Place Success Rate: 50%
• Multi-step Long Task Success Rate: 62%

The key point is: the core model was not altered. This is important for actual deployment—retraining completely is both expensive and slow.

Why the Robotic Arm Model Cannot Fly Directly

VLA can transfer well in “understanding the scene + comprehending the task,” but controlling dynamics cannot be directly transferred:

• Robotic arm data operates in a mostly stationary environment
• Drones are underactuated systems, and error accumulates quickly, leading to crashes if not careful
• The physical laws and control constraints on both sides are fundamentally different

How They Solved It

Two core ideas:

1. Add Physical Constraints During Inference: Instead of embedding new dynamics into the model, correct them online according to physical laws at the output stage.
2. Use Gaussian Splatting to Create Navigation Data: Avoid the need to run around the world collecting data with real machines.

This approach of “adding modules to the base model without end-to-end retraining” aligns with AIR-VLA and DroneVLA, but takes a different angle.

Who Will Benefit from This

Companies involved in aerial operations (logistics, inspections, search and rescue) may find this interesting:

• No need to gather a large amount of drone data
• The hybrid approach of physical guidance + AI is more controllable in safety-sensitive scenarios, unlike purely learning-based control which can be quite mystical.

How to View This Matter

Conclusion: This direction is still relatively early-stage. The most relevant teams are those engaged in aerial operations—robotics/drone manufacturers, research laboratories, and solution providers. Short-term trading is of little significance, but long-term investors can pay attention to key milestones from research to scaling.