Learning Energy-Efficient Air--Ground Actuation for Hybrid Robots on Stair-Like Terrain

A research team from Carnegie Mellon University and the University of Hong Kong has published a paper detailing a novel reinforcement learning (RL) framework that teaches hybrid robots to navigate stair-like obstacles with dramatically improved energy efficiency. The core innovation is a single, continuous AI policy that learns to seamlessly coordinate a robot's wheels, propellers, and tilt servos in real-time, eliminating the need for pre-programmed switches between 'ground' and 'flight' modes. Trained in NVIDIA's Isaac Lab simulator using hardware-calibrated power models, the policy's reward function directly penalizes true electrical energy consumption, pushing it to discover efficient 'thrust-assisted driving' behaviors.

In simulation, this learned policy used about 4 times less energy than a controller relying solely on propellers. The team successfully transferred the policy to a physical robot prototype named 'DoubleBee.' On a real-world task of climbing an 8cm gap—a proxy for a stair edge—the AI-driven system achieved a 38% lower average power draw compared to a traditional, rule-based controller that operates the aerial and ground systems separately. This demonstrates that energy-efficient hybrid locomotion can emerge from end-to-end learning and be deployed on actual hardware, a significant step toward more practical and enduring robots for complex environments like disaster zones or construction sites.