--- license: other task_categories: - robotics tags: - mujoco - escape-room - mobile-manipulation - imitation-learning pretty_name: Escape-room test set (150 two-robot escape houses) --- # Escape-room test set: 150 two-robot escape houses with expert demonstrations Procedurally generated multi-room escape houses for a simulated Stanford TidyBot (holonomic base + Kinova Gen3 + Robotiq 2F-85) in MuJoCo (`escape-room-sim`: `tools/gen_escape_houses.py`, `tools/solve_gen_houses.py`, `tools/build_test_set.py`). Every house here was solved by the scripted expert; 150 were sampled from the 250 solved of 260 generated houses. Two TidyBots (`tidybot`, `tidybot_2`) per house. Modes: **hold** (a door needs two buttons held at the same time, one robot on each), **split** (the robots start in different rooms; each keypad's code is on a whiteboard in the *other* robot's room, so they must share it), **relay** (each robot's puzzle unlocks the other robot's door), **pair** (one room, two different puzzles, one per robot, that both unlock the exit). Puzzles: red button (press), keypad (type the 4-digit code written on a whiteboard: `Code: `), wall lever (push it down). Whiteboards hang at 1.33-1.78 m, readable from the robot's head-camera height up close. Goal: the exit door opened. * rooms: 50 x 1, 50 x 2, 50 x 3 * modes: hold 52, pair 30, relay 35, split 33 * puzzles: button 157, keypad 120, lever 70 * 104,649 demo steps at 10 Hz; episodes 34-166 s of simulated time (median 61 s) ## Files | path | content | |---|---| | `index.jsonl` | one line per episode: name, paths, rooms, room shapes, puzzles, robots, mode, turns, steps, sim seconds | | `episodes//demo.npz` | the demonstration | | `tasks/gen/.task` | the task graph (its `house:` path is relative to the repo root) | | `houses/gen/.json` / `.txt` / `.png` | the house, its ASCII floor plan, a top-down preview | To simulate a house, put `houses/` and `tasks/` into an `escape-room-sim` checkout (same relative paths). ## Demonstrations (`demo.npz`) 10 Hz, recorded through the robots' own actuators (`agent/demos.py`). Keys: `r1_image_base`, `r1_image_wrist`, `r1_state`, `r1_ee_pos`, `r1_ee_quat`, `r1_action` and the same with `r2_` (one set per robot), `robots`, `mode`, `turns`; `t`, `instruction`, `task`, `hz`, `success`, `node_states`, `log`. Images `[T, 224, 224, 3]` uint8; `state` `[T, 11]` (base x, y, yaw, 7 arm joints, gripper); `action` `[T, 11]` joint-space setpoint 0.1 s later. The demonstrator is a **privileged scripted expert**: it reads the simulator state and the task file, including each keypad's code; it never reads the whiteboards. Its two robots follow a fixed schedule (no messages). ## Licences Rendered images show THOR models from MolmoSpaces (`allenai/molmospaces`, CC BY 4.0) and the TidyBot model from MuJoCo Menagerie (Apache-2.0). Puzzles and signs are procedural (escape-room-sim).