PocketSPAR

PocketSPAR is a research app for Stable Planning through Aligned Representations (SPAR) in Model-Based RL for planning with learned heuristics that runs fully on-device.

PocketSPAR is a research app for studying visual planning with Stable Planning through Aligned Representations (SPAR) on iPhone and iPad.

The app runs prepared PocketSPAR profiles. A profile can include Core ML models, observation specifications, action labels, latent dimensions, and search settings prepared with PocketSPAR Bundler.

PocketSPAR supports SPAR-compatible domains such as Sokoban and Rubik's Cube. Profiles define their own required view count, action space, latent dimension, and search configuration.

Core workflows:

• Install a .pocketsparprofile file or a trusted HTTPS catalog profile.
• Capture or import start and goal observations.
• Inspect decoded latent-state reconstructions before search.
• Configure compute policy, memory strategy, batch size, and node limits.
• Run local Q* search with the profile's world model and heuristic function.
• Review solution cost, generated nodes, expanded nodes, elapsed time, and action sequence.
• Save sessions and export outputs for later comparison.

Camera frames and imported images are processed on device unless you explicitly export or share them, or send a solution preview to a local-network SokobanDemo instance.