USTB team redesigns robotic butterfly wings via deep reinforcement learning

Researchers at USTB and partner labs showed that a deep reinforcement learning loop fed by real flights, not just simulation, can redesign the wings of a biomimetic flapping robot. The 48.3-gram USTButterfly-II spans 57.5 cm and flaps at 5 Hz with an 84-degree stroke; after 95 flights, a lift-focused design raised the lift parameter from 2.23 to 9.8 (+339%), while a thrust-focused one lifted average speed from 1.98 to 2.92 m/s.

• USTButterfly-II: 48.3 g, 57.5 cm span, 5 Hz, 84-degree stroke
• 95 real flights across 19 wing-shape pairs, tracked by 14 Optitrack cameras
• Lift parameter rose from 2.23 to 9.8 (+339%), mean climb angle 32.7 degrees
• Thrust-focused design raised average speed from 1.98 to 2.92 m/s