RoboQuest Demonstrations

Generalist Physical Agents that Search, Inspect and Test

Project Page | Paper (coming soon) | Code

RoboQuest is a benchmark for goal-directed embodied exploration, in which a mobile manipulator must gather task-relevant information through physical interaction, act on the evidence, and decide for itself when the task is done. This repository holds its demonstration dataset: 5,000 verified demonstrations of the ten RoboQuest tasks (500 per task, 366 hours at 20 Hz) by a Franka Panda arm on a mobile base in RoboCasa365 kitchens, each with a two-level language annotation (stages and subtasks).

Contents

Path	Format	What it is
lerobot/<task>/	LeRobot v2.1	Three camera streams as video, the robot state and action at every frame, the goal text as the task, and per-frame stage and subtask indices
raw/<task>/<id>/	simulator recordings	The simulator state at every tick, the actions, the stage segments and the captions: enough to replay an episode exactly and render it again at any resolution
raw/manifest.json	JSON	Every episode with its kitchen layout and style, task factors, length and the checksum of its states file

The scene of every demonstration (suite/v1/<task>/demos/) and the tools that produced the LeRobot datasets (scripts/dataset/) are in the code repository. The demonstration scenes are development scenes; none of them is one of the benchmark's 500 evaluation instances.

Tasks

Family	Task	Episodes	Frames	Hours
Search & Explore	Locked Storage (locked_storage)	500	2,238,904	31.1
	Search Room (search_room)	500	1,807,513	25.1
	Blackout Search (blackout_search)	500	4,471,462	62.1
Object Inspect	Painted Cubes (painted_cubes)	500	3,053,005	42.4
	Marked Mugs (marked_mugs)	500	2,018,596	28.0
	Unfamiliar Containers (unfamiliar_containers)	500	3,813,854	53.0
Testing	Puzzle Box (puzzle_box)	500	983,469	13.7
	Stamp Composition (stamps)	500	2,165,174	30.1
	Wobbly Stand (wobbly_stand)	500	1,671,246	23.2
	Odd Parcel (odd_parcel)	500	4,159,509	57.8
Total		5,000	26,382,732	366.4

LeRobot features

Each lerobot/<task>/ is a LeRobot v2.1 dataset (robot_type: panda_omron, 20 fps):

Feature	Shape	Contents
image, right_image	256 × 256 × 3 video	the left and right scene cameras
wrist_image	256 × 256 × 3 video	the camera on the gripper
state	16	end-effector position (3) and quaternion xyzw (4) in the robot base frame, base position (3) and quaternion xyzw (4) in the world, the two gripper finger joints (2)
actions	12	the native RoboCasa PandaOmron action: arm motion (6, controller deltas in the base frame), gripper (1: > 0 closes, < 0 opens), base velocity (3), torso (1), mode (1)
stage_index, subtask_index	1	the frame's stage and subtask, indexing meta/stages.jsonl and meta/subtasks.jsonl
task_index	1	the goal text, in meta/tasks.jsonl

Usage

# one task as a LeRobot dataset
from huggingface_hub import snapshot_download
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

root = snapshot_download("declare-lab/RoboQuest", repo_type="dataset", allow_patterns="lerobot/puzzle_box/*")
dataset = LeRobotDataset("roboquest/puzzle_box", root=f"{root}/lerobot/puzzle_box")
frame = dataset[0]

# the raw recordings of one task, then render an episode again at another resolution with the code repository
hf download declare-lab/RoboQuest --repo-type dataset --include "raw/puzzle_box/*" --local-dir roboquest-data
git clone https://github.com/declare-lab/RoboQuest && cd RoboQuest
MUJOCO_GL=egl python scripts/dataset/rerender.py ../roboquest-data/raw/puzzle_box/<id> out --sim-size 512 --out-size 256 --mp4

Citation

@misc{liu2026roboquest,
  title  = {{RoboQuest}: Generalist Physical Agents that Search, Inspect and Test},
  author = {Liu, Renhang and Majumder, Navonil and Pala, Tej Deep and Poria, Soujanya},
  year   = {2026},
  note   = {arXiv preprint (coming soon)}
}

License

To be announced.