01 / 2026In progress
Six-panel SmartFly CNN–RNN stable-flight recording with external cameras, trajectory, actual eye inputs, height, and speed.
Recorded stable flight · A fixed-seed example, not a forest-navigation result.

SmartFly

Fly-inspired multimodal embodied navigation

A dual-task navigation framework in a MuJoCo fly simulation, combining visual obstacle navigation and vision–odor source seeking, with CNN–RNN and FlyVis–RNN encoder variants.

The project spans hierarchical PPO control, forest and dynamic odor environments, curriculum training, and representation analysis. CNN–RNN stabilization replays are available in both observation modes; the full navigation comparison is ongoing.

Embodied AIPPOMuJoCoPyTorchFlyVis
Read project overview ↗
Explore my GitHub ↗