Sokoban
A puzzle task environment in which an agent pushes boxes to designated goal positions. It is used to evaluate an agent's long-horizon planning and spatial reasoning abilities.
Timeline (last 90 days)
Agima was validated in four environments—Maze2D, Sokoban, Pacman and PushT—and can generalize to harder settings unseen in training.
After SFT training only on the Sokoban "two-box" version, Agima solved 40% of "four-box" puzzles, and it saw the real board only once every five steps.
On the four-box Sokoban task, an action-only planner solved only 2%, far below Agima's 40%.
Agima has the same omni-model serve simultaneously as both policy and world model, so that imagination and decision-making share representations, which is the key to its ability to complete four-box Sokoban with sparse real feedback.
ViSkill achieves an overall success rate of 0.89 on the three environments Sokoban, FrozenLake, and PrimitiveSkill, rising to 0.91 when cold start is enabled.
Sokoban tests long-horizon planning and spatial deduction, FrozenLake focuses on decision-making to avoid traps in stochastic environments, and PrimitiveSkill focuses on the reuse of basic skills.
All Facts (6)
Agima was validated in four environments—Maze2D, Sokoban, Pacman and PushT—and can generalize to harder settings unseen in training.
50%UnverifiedAfter SFT training only on the Sokoban "two-box" version, Agima solved 40% of "four-box" puzzles, and it saw the real board only once every five steps.
50%UnverifiedOn the four-box Sokoban task, an action-only planner solved only 2%, far below Agima's 40%.
50%UnverifiedAgima has the same omni-model serve simultaneously as both policy and world model, so that imagination and decision-making share representations, which is the key to its ability to complete four-box Sokoban with sparse real feedback.
50%UnverifiedViSkill achieves an overall success rate of 0.89 on the three environments Sokoban, FrozenLake, and PrimitiveSkill, rising to 0.91 when cold start is enabled.
50%UnverifiedSokoban tests long-horizon planning and spatial deduction, FrozenLake focuses on decision-making to avoid traps in stochastic environments, and PrimitiveSkill focuses on the reuse of basic skills.
50%