Self-Play Distillation
A technique in which a student policy continuously generates trajectories through interaction with a world model and improves itself, compressing planning-level behavior into a feedforward network; the distilled goal-conditioned policy can output actions directly without online optimization.
Timeline (last 90 days)
The DSP research team proposes distilling a closed-loop, goal-conditioned policy inside the world model through self-play, compressing planning capability into a fast policy network to reduce the computational overhead of online planning.
The distilled goal-conditioned policy takes the goal state as an input condition and directly outputs actions without online optimization solving, greatly improving inference speed, but at the cost of sacrificing some planning precision.
All Facts (2)
The DSP research team proposes distilling a closed-loop, goal-conditioned policy inside the world model through self-play, compressing planning capability into a fast policy network to reduce the computational overhead of online planning.
50%UnverifiedThe distilled goal-conditioned policy takes the goal state as an input condition and directly outputs actions without online optimization solving, greatly improving inference speed, but at the cost of sacrificing some planning precision.
50%