FaSt World PlANner: Advantage-Aware
Action-Chunk Policy for Online Planning

Explore demonstrations
01Planning mechanics

k-MPPI
Visualization

MPPI uses one global search. k-MPPI combines local refinement and global exploration to maintain multiple search modes.

Open visualization video
02Unitree G1 + Go2

Simulation
Evaluation

SWAN is compared with BMPC on figure-eight and straight-line velocity tracking for simulated Unitree Go2 and G1 robots.

SWAN

Tracking Evaluation

SWAN vs. BMPC: commanded and simulated body velocities during figure-eight and straight-line locomotion.

Unitree Go2

Figure-8 Tracking

SWAN (Ours)
BMPC (best baseline)

Straight-Line Tracking

SWAN (Ours)
BMPC (best baseline)

Unitree G1

Figure-8 Tracking

SWAN (Ours)
BMPC (best baseline)

Straight-Line Tracking

SWAN (Ours)
BMPC (best baseline)
03Unitree Go2

Real-World
Deployment

SWAN controls a real Unitree Go2 across directional locomotion, turning, and external pushes.

Tracking Evaluation

Commanded and motion-capture velocities during figure-eight and straight-line locomotion.

Figure-8 Tracking

Straight-Line Tracking

k-MPPI Visualization
1 / 6

Initial