justintiensmith/MT_State_Recognition_New

MT State Recognition New

This is the curated 200-episode state-recognition fine-tuning dataset used for the missing-cell compositional-transfer experiment. It was assembled with LeRobot v3.0 at 30 FPS and contains five RGB camera streams, six-dimensional robot state, and six-dimensional actions.

Each episode's task_index maps to its final training prompt in meta/tasks.parquet. The episode-level provenance, final prompt, scene description, target object, training role, source repository, and source episode are recorded in meta/merge_manifest.parquet and meta/merge_manifest.csv.

Composition

• justintiensmith/MT_State_Recognition_200_OLD: 134 selected episodes

• justintiensmith/MT_State_Recognition_20: 20 selected episodes

• justintiensmith/SP_State_Recognition_61: 28 selected episodes

• justintiensmith/MT_State_Recognition_40: 18 selected episodes

• Total episodes: 200

• Total frames: 100378

• Unique training prompts: 48

Held-out object-state combinations

• upside-down white cup
• open yellow-white striped cup
• closed brown cup
• upright glass spice container

These combinations are excluded from positive state-qualified training prompts. State-masked demonstrations can still expose the corresponding objects and physical states without naming the held-out state.

Visualize