Embodied intelligence study / experiment 01

Fly’s Gambit

Can a neural navigator steer a physically simulated fly across a chessboard? Watch the verified run.

MuJoCo / physical rollout
Orange cylinders are obstacles. Cyan marks F7. The glow follows the simulated fly. 64 frames / A1 → F7
Why this matters

See. Choose. Move.

Embodied intelligence is easier to understand when its jobs are separated and inspectable.

Overhead MuJoCo camera view of the obstacle board
01 / See

A trained camera model turns rendered pixels into occupied squares.

02 / Choose

The trained navigator turns the board, start, and target into one of eight directions.

03 / Move

A published FlyBody controller turns that direction into joint and adhesion control.

Interrogate layer 02

Change the target. Watch the planner choose.

This board runs the saved navigation network in your browser. The default A1 → F7 route is the command used in the physical simulation above.

Choose a destination step 00