WatchAct: A Benchmark for Behavior-Grounded Robot Manipulation

🌐 Project Page  |  🤗 Dataset  |  💻 Code

WatchAct is a behavior-grounded benchmark for robotic manipulation, where the robot reasons over observed human behavior and a language instruction to perform the corresponding task.

Files

.
├── data/<task>.jsonl          # evaluation instances (one row per video × view × variant)
├── meta_data/<task>.json      # ground-truth scene definition, initial states, and goal
├── videos/<task>/             # multi-view recordings: front/, left/, left_back/ (*.MP4)
└── bddl_files/<task>/         # LIBERO task definition files (*.bddl)

• data/ — the model-facing evaluation rows. Each row provides a language_instruction, the video to watch, its camera_perspective / spatial_reference, and links back to its meta_data example via original_id and to its task file via bddl_file.
• meta_data/ — per-example ground truth: the object list, symbolic initial_states, the target final_goal, and the oracle_plan used for scoring.
• videos/ — three synchronized views (front, left, left_back) of a person performing each task.
• bddl_files/ — the corresponding LIBERO simulation task definitions.

A <task> is the file stem (e.g. Ordinal, Fine-Grained_Action). The 14 tasks, grouped by capability:

Category	Tasks
Procedural Reasoning	Imitation, Reversal, Temporal_Sort
Event Grounding	Fine-Grained_Action, Count, Ordinal, State_Change, Moment
Implicit Intent Inference	Nonverbal_Cue, Reference_Disambiguation
Episodic Reasoning	Restore_Previous_State, Task_Continuation, Error_Correction, Conditional_Execution

See the code repository for the full evaluation pipeline.