image imagewidth (px) 1.13k 1.13k |
|---|
Agentic Robotics Last Exam: task assets
This repository holds the large files of the tasks in the Agentic Robotics Benchmark: photo sets, document scans, meshes and other inputs that are too large to keep in the GitHub repository. Each task's definition, grader and run instructions stay in GitHub. This repository only serves the bytes.
Layout
The tree mirrors the benchmark's tasks/ directory, so every task has its own folder:
tasks/<platform>/<task name>/
payload/ files the agent receives; the same relative paths as the task's setup/payload/
sources/ (optional) authoring inputs used to regenerate the payload
shared/ (optional) assets used by several tasks
<platform> is one of the benchmark's platform keys (arm, hand, quadruped, humanoid, uav, ground,
underwater, space, soft, surgical, swarm, subsystems). <task name> is the task's name in its
task.yaml.
How a task uses it
- A task never follows a branch of this repository. Its
image/assets.sha256names a commit of this repository and lists the sha256 of every file it needs. image/fetch_assets.pydownloads those files at that commit while the image is built and fails the build on any mismatch.- Later uploads for other tasks therefore never change a task that is already published.
To add a task's assets, upload its folder under tasks/<platform>/<task name>/. Then record the resulting commit and
the file hashes in the task's image/assets.sha256 in the same pull request that adds or updates the task
(scripts/task_assets.py list in the benchmark writes it).
The benchmark names this repository in one file, task-assets.toml at its root, and copies that name into every task's
assets.sha256 (scripts/task_assets.py sync). If this repository is renamed or mirrored, the benchmark changes that
file and runs the sync; see the benchmark's docs/task-assets.md.
Tasks
| task | platform | contents |
|---|---|---|
robotic_arm_requirements_replica_from_photos |
arm | 31 survey photos of a desktop arm with photos.json, 30 document pages (servo, bus-board and PSU datasheets and an erratum sheet), 2 servo-lot inspection reports (scanned images) |
robotic_arm_cad_design_from_photos |
arm | 80 photos of a modified six-servo desktop arm (teardown of the ten printed parts on a cutting mat with scale references, close-ups, the re-assembled unit in three poses, the gripper; Blender Cycles in a lab room), the servo's datasheet and dimensioned drawing (images), its vendor STEP and keep-out STEP |
robotic_arm_full_assembly_so101_panda |
arm | the 25-photo picture manual of a modified six-servo desktop arm with photos.json, the ten printed parts' STL files and the servo's CAD (STL), the Franka Panda MuJoCo model (Apache-2.0, from MuJoCo Menagerie, its LICENSE included); payload/render/: the scenery of the agent cameras' renderer (the factory hall room/, built from the FactoryWarehouse Interior Modular Pack and the Archviz Laboratory Pack, Unity Asset Store, with its Blender Cycles lightmaps and image light; Poly Haven's oak_veneer_01 maps oak/, CC0); sources/render/: an earlier bake of the factory room (v2), used for the maintainers' videos only |
robotic_arm_full_assembly_openarm_panda |
arm | the 27-photo picture manual of an OpenArm 2.0 pair with photos.json (Blender Cycles), the 18 loose units' looks (given/parts/, STL) and the convex pieces the cell simulates for them (given/collision.tar, unpacked at image build into given/collision/<unit>/), cut from the OpenArm 2.0 assembly STEP (Enactic, CERN-OHL-S-2.0); the Franka Panda MuJoCo model (Apache-2.0, from MuJoCo Menagerie, its LICENSE included); payload/render/: the scenery of the agent cameras' renderer, the same files as robotic_arm_full_assembly_so101_panda's |
quadruped_cad_design_from_photos |
quadruped | 132 photos of a modified 12-joint quadruped (teardown of the 42 printed parts on an A2 cutting mat in seven groups with scale references, 20 close-ups, the re-assembled robot in three configurations, two legs in front of a checker board; Blender Cycles in a lab room), the actuator unit's and the foot pad's datasheets and dimensioned drawings (images), their vendor STEP files and the actuator's keep-out STEP; NOTICE.md next to payload/: the open-hardware design the robot and the two components derive from; sources/render/: the authoring inputs the photos were rendered with, not fetched by the task (the lab room, rebuilt from the Unity Asset Store pack "Archviz Laboratory Pack" 2.0, as an arb-room/1 bundle; Poly Haven's schadowplatz HDRI, CC0) |
arm_hidden_mechanism_box |
arm | payload/render/ibl.ktx: the image light of the agent cameras' renderer (MuJoCo's Filament), a light probe of the task's LIBERO kitchen scene (lerobot/libero-assets at 0b3ea86b) path traced with OVRTX and converted to a KTX cube map with Filament's cmgen; made for this task |
bimanual_cad_design_from_photos |
arm | 153 photos of a modified two-arm unit (teardown of the 29 made parts on an A3 cutting mat with scale references, both sides, 43 close-ups, the re-assembled unit in three configurations, both grippers closed and open; Blender Cycles in a lab room), the bought parts' datasheets and dimensioned drawings (images), their vendor STEP and keep-out STEP files (four actuator types, two gripper modules, the column stand) |
quadruped_requirements_replica_from_photos |
quadruped | 36 survey photos of a small 12-actuator, belt-reduced quadruped on a lab bench with photos.json (Blender Cycles in a lab room), 35 document pages (actuator-module, battery-pack, driver and hub datasheets and three vendor erratum sheets), 3 incoming-inspection reports (two actuator-module lots and one battery cell lot; scanned images); sources/render/: the maintainers' render assets of the photo set (not fetched by the task): the lab room bundle built from the Archviz Laboratory Pack (Unity Asset Store), Poly Haven's Schadowplatz HDRI and oak_veneer_01 maps (CC0) |
robocasa_object_dynamics_from_video |
arm | payload/data/: the agent's input recordings of a Franka Panda poking four household objects (a practice object and three to identify) with a ball-tipped rod in RoboCasa kitchens, 40 clips of three calibrated cameras (120 MP4 videos, rendered with MuJoCo's Filament under an image light distilled from an OVRTX path-traced render) and per clip the commanded ball-tip path at 500 Hz and the arm's joint angles per frame (80 NPY arrays); the objects, kitchen textures and clutter are RoboCasa assets (CC BY 4.0, Nasiriany et al., RoboCasa, RSS 2024), the arm robosuite 1.5.2's Panda (MIT); attribution in the task's verify/NOTICE.md in the benchmark repository |
hot3d_sharpa_retarget |
hand | nine HOT3D Aria clips in which a person moves one object with both hands, under clips/<clip>/: the RGB frames (lossless PNG, original fisheye orientation), frames.jsonl (60 Hz control step to frame), the per-frame MPS camera calibration, the dataset's calibration snapshot and the headset pose in the simulator frame; derived from the HOT3D dataset (Meta), CC BY-SA 4.0, its LICENSE-HOT3D.txt next to payload/ |
visual_low_level_flight_control |
uav | payload/render/: the scenery of the onboard camera's renderer (MuJoCo's Filament): the main hall of the Unity Asset Store pack "Warehouse - Abandoned Factory District 1.1" (ScansFactory), cleared around the flight volume, as meshes per material with 256 px textures (warehouse/); the course's worn-paint textures and floor paint (course/, made for this task); the image light ibl.ktx and sky skybox.ktx, cube maps of an OVRTX path-traced light probe of the hall, made with Filament's cmgen for this task |
dexhand_shape_sorter_all_pieces |
hand | the agent's images: per scene (instances/inst_a..d, practice/practice_1..3) two calibrated RGB-D cameras' start images and the 16 images of a hand-eye calibration routine (140 PNGs), the scene in an office of the Unity Asset Store pack "Office With Conference Room", rendered with MuJoCo's Filament under an image light distilled from an OVRTX path-traced light probe (made for this task); payload/superdex_assets/: the public SuperDex assets the task uses, the shape_box prefab (CC BY 4.0) and the Wuji Hand 2 right hand (MIT, its NOTICE), each folder with its LICENSE |
furniture_assembly_from_drawings |
humanoid | payload/render/room_decor/: the room's decor as the agent cameras' renderer (MuJoCo's Filament) draws it; the image keeps it in a root-owned folder, it is not staged into the agent home. The furnished room of the task (the oak floor, the plaster walls and ceiling, the skirting, the window's frame, glass and board, the door leaf, the pictures, the clock, the ceiling panels, the couch, the plants, the rug and the buildings across the street) as meshes (NPZ) with base maps of at most 512 px and decor.json (the meshes, their materials and base maps), made from the Unity Asset Store packs "Office and Police Station Pack" and "Office With Conference Room"; the image light ibl.ktx, the mean of three OVRTX path-traced light probes of the room with the sunlit patches and the lamps' emissive surfaces painted out, prefiltered with Filament's cmgen; made for this task |
bimanual_requirements_replica_from_photos |
arm | 41 survey photos of a two-arm robot (two seven-joint arms with two-finger grippers on a pedestal column) on a lab bench with photos.json, 44 document pages (joint-module, bus-board and power-supply datasheets and the module vendor's erratum sheet), 2 module-lot inspection reports (scanned images) |
robotic_arm_robot_description_from_photos |
arm | 21 photos of a modified six-servo desktop arm on a lab bench with photos.json, README.md and the servo bus log of every photo (six poses from three sides each, two gripper close-ups, a base close-up; path traced with NVIDIA OVRTX in a seamless grey photo cove), the published printed parts, the published base in the world frame and the servo (STL, parts/ with parts.json), the servo's datasheet page and dimensioned drawing (images), the deliverable contract's two figures (contract/) |
fault_adaptive_spacecraft_docking |
space | payload/render/: the docking camera's scene as the grader's renderer (MuJoCo's Filament) draws it, kept root-only in the image: the target drum (NASA Wind) with its boom, drogue and LED docking targets, the chaser (NASA 2U CubeSat), 13 textured materials (multi-layer insulation, cells, aluminium, paint; normal and roughness maps; some made from the owner's ISS pack textures), the image light ibl.ktx distilled from an OVRTX path-traced light probe; payload/given/frames/: 40 sample docking-camera frames from development episodes with their geometry, for the agent; NASA 3D Resources models (public domain); made for this task |
driving_safety_supervisor |
ground | the image's build-time downloads, mirrored byte for byte from their sources and fetched by the image build, not staged into the agent home: carla/CARLA_0.9.15.tar.gz, the CARLA 0.9.15 Linux release of downloads.carlasim.com (unpacked into /opt/carla/server); models/: the Hugging Face snapshots of openai/clip-vit-base-patch16 (57c21647) and google/owlv2-base-patch16-ensemble (cfd3195b), the files the image keeps in /opt/models; torchvision/: torchvision's COCO Faster R-CNN ResNet-50 FPN v2 weights of download.pytorch.org (fasterrcnn_resnet50_fpn_v2_coco-dd69338a.pth, re-saved by the image build) |
aerial_search_geolocation |
uav | the images' build-time downloads, mirrored byte for byte from their sources and fetched by the image builds, not staged into the agent home: scenes/: the five parts of the closeloop_envs archive of the Hugging Face dataset wangxiangyu0814/TravelUAV_env (44de5739), TravelUAV's packaged Unreal Engine scenes with the AirSim plugin (the agent's image unpacks the two practice scenes, the separate verifier image the three held-out ones); spawn_areas/map_spawnarea_info.json: AeroVLA's per-map placement areas (github.com/XuPeng23/AeroVLA at 2d7ebd6); models/: the CLIP and OWLv2 snapshots, the same files as driving_safety_supervisor's |
hazard_bay_build |
swarm | the 12 photographs of the bay the agent rebuilds (given/photos/P01-P12.jpg, 1600 × 1200, path traced with Isaac Sim 6.2's RTX renderer; their cameras.json stays in the benchmark repository); payload/arb/: art of the task's own asset library, visual only (no physics), moved by the image build into /opt/assets/arb/ and not staged into the agent home: the plant's hall round the bay (arb/hall/: bay_hall.usda, its prototypes, 1K-2K textures, hall_manifest.json naming each prototype's source prefab), built for this task from the Unity Asset Store packs "FactoryWarehouse Interior Modular Pack", "Factory Interior Factory Props Vol 1" and "Warehouse Props Vol 1" (CyberneticWalrus), "Warehouse - Abandoned Factory District" (ScansFactory) and "Military Base ... Military Props" (FreshCan3D); the four airframes' look layer (arb/airframes/look/), the parts of the industrial hexacopter of the Unity Asset Store pack "Realistic Drones" fitted to the task's four rotor layouts; made for this task |
novel_object_6dof_grasp_detection |
arm | a derived subset of GraspNet-1Billion (real RealSense RGB-D captures of cluttered tabletops), as two archives with their manifests and the NOTICE: payload/graspnet/: the image's build-time data, unpacked into the root-owned /data/graspnet and not staged into the agent home (20 development scenes and 12 practice scenes, 8 viewpoints each, and the grasp annotations of ten objects; the image build then removes the meshes and most object poses); verifier/: the frames the verify stage fetches (28 novel-object scenes, 16 viewpoints each, with the poses and meshes of their 26 objects that the scoring needs), never in an image |
uav_requirements_replica_from_photos |
uav | 12 survey photos of a guarded indoor quadrotor in a workshop, of two dated build stages (1536 × 1152 lab exports of Blender Cycles renders; the scene is procedural, made for this task), 8 vendor datasheet pages and 2 incoming-lot inspection reports (scanned images); the vehicle's guarded core derives from AgiPix (github.com/SasaKuruppuarachchi/agipix at 147a736b, CC0-1.0), the documents are made for this task |
uav_cad_design_from_photos |
uav | 155 photos of an AGIPIX v2 research quadrotor (teardown of its 51 made parts in 24 families on A3 and A2 cutting mats with scale references, both sides; 31 close-ups; the assembled unit and a 22-photo disassembly sequence; Blender Cycles in a lab room), and for each of the 15 purchased component types (motors, propellers of both hands, Livox MID-360, Intel RealSense D455, the Jetson Orin NX stack, Cube Orange, ESC, power board, battery, connectors, standoffs, dampers) a STEP of its original geometry in a documented frame, a datasheet and a dimensioned drawing; the design is AGIPIX v2 (github.com/SasaKuruppuarachchi/agipix at 147a736b and the Autodesk share its cad/README.md links, CC0-1.0), unchanged |
uav_assembly_agipix_xarm7 |
uav | two archives that the image build unpacks into the agent's givens: payload.tar.gz: the adapted AGiPix quadrotor kit (given/kit/: the source meshes of its 24 parts of 15 kinds, the convex pieces the cell simulates for them, kit.json and the provenance records; derived from github.com/SasaKuruppuarachchi/agipix at 147a736b, CC0-1.0, its licence included), the UFACTORY xArm 7 MuJoCo model with its gripper (given/xarm7/, BSD-3-Clause, from MuJoCo Menagerie, its LICENSE included) and the scenery of the agent cameras' renderer (render/, the same files as robotic_arm_full_assembly_so101_panda's); manual.tar.gz: the 27-page picture manual of the adapted kit (given/manual/, orthographic diagrams of the supplied meshes), unpacked over the copy inside payload.tar.gz |
robotic_arm_environment_twin_from_real_photo |
arm | one real colour photo, photo.png (1920 × 1080): image 000000 of validation scene 000009 of the HOPE dataset (NVIDIA; Tyree et al., arXiv:2203.05701), an Intel RealSense D415 capture of a console table with 19 toy grocery items on its top, in its pulled-out drawer and on a shelf; byte for byte from the BOP distribution (Hugging Face dataset bop-benchmark/hope at ddd0a26c). The task's ground truth (the dataset's meshes and poses as a true twin) stays in the benchmark repository |
house_cleaning_with_tools |
humanoid | payload/given/cell/survey/: the survey photos of the studio apartment, two development scenes (scene_a, scene_b; drawn by the task's rules, not among the graded scenes) as the task's own cell and cameras deliver them (MuJoCo's Filament): per scene the three ceiling RGB-D cameras and six head-camera views, each frame as its reply's metadata (JSON), colour (PNG) and 16-bit depth in millimetres (PNG), and a README.md; the image keeps them root-owned in /opt/arb-assets/survey and the setup stages a copy into the agent home. The room's decor is furniture_assembly_from_drawings's (the same apartment); made for this task |
quadruped_full_assembly_12dof_ur10e |
quadruped | payload/: the givens of a twelve-actuator quadruped that a UR10e-class arm assembles: the 25-photo picture manual with photos.json, the six unit types' STL files, the actuator module's datasheet and drawing (PNG) and STEP files, the arm and gripper MuJoCo models (BSD-3-Clause and BSD-2-Clause, their LICENSE files next to them) and render/, the agent cameras' scenery (the baked factory hall, the bench top's wood maps); verifier/gt.tar.xz: the grader's kit of the cell (kit.json, the units' meshes and carved collision pieces), downloaded by the verify stage, never in an image; sources/render/: the hall's authoring inputs (unbaked, the bake with its probe, the HDRI) |
quadruped_environment_twin_from_real_photos |
quadruped | 44 real colour photos, photos/site_001.jpg .. site_044.jpg (6208 × 4135): the 44 undistorted DSLR images of the high-resolution multi-view training scene delivery_area of the ETH3D benchmark (Schöps et al., CVPR 2017, https://www.eth3d.net/), an underground goods delivery bay photographed from a tripod; byte for byte from delivery_area_dslr_undistorted.7z, renamed in capture order (DSC_0675 .. DSC_0718). The task's ground truth (a true twin built from the dataset's laser scan and camera calibration) stays in the benchmark repository |
quadruped_hazardous_plant_inspection |
quadruped | payload/render/qinspect_look_v1.tar: the look of the task's plant buildings as the robot cameras' renderer (MuJoCo's classic renderer) draws them; the image unpacks it into a root-owned folder that the agent's account cannot read, it is not staged into the agent home. One uncompressed tar of 62 files and a manifest.json that lists every file with its sha256 and every item with its pack and source: 21 surface textures (base-colour maps of 512 to 1024 px for walls, floors, ceilings and plant items), 10 props (two lamps, a cable tray, a barrel, a pallet, a shipping crate, a fire alarm, a fire extinguisher, an electric box and an electric motor) as OBJ meshes with their base-colour maps, 12 safety signs cut from a pack's decal sheet, and 4 small effect maps generated by the bundle's builder; made for this task from the Unity packs "Chernobyl" 1.1 and "Nuclear power plant" 1.0 (VP.Studio), KitBash3D's "Refineries" kit (4K), and the Unity Asset Store packs "Factory Interior + Props Vol 1" and "Warehouse Props Vol 1" (CyberneticWalrus) and "Warehouse - Abandoned Factory District" 1.1 (ScansFactory). The buildings (site packages), instruments, nameplates and drawings are generated by the task and stay in the benchmark repository |
cross_embodiment_adaptation |
quadruped | payload/visual_assets/: what the task's robot service renders its two cameras with; the image moves it into a root-owned folder that the agent's account cannot read, it is not staged into the agent home. 114 files: the Unitree Go1 and Bitcraze Crazyflie 2 models of MuJoCo Menagerie (MJCF, meshes, their LICENSE files), twelve Poly Haven texture maps (PNG), three Poly Haven models baked to OBJ parts with colour, normal and roughness maps (baked/), and provenance.json with every file's source URL and sha256. |
racecar_requirements_replica_from_photos |
ground | the images among the givens of the requirements task in which the agent specifies six replicas of a tenth-scale autonomous race car that can no longer be measured: 33 photographs of the reference unit (photos/PH-01.jpg .. PH-33.jpg, 2016 × 1512 JPEG; renders of the benchmark maintainer's own model of the car in a lab room) and 27 page images of the catalogue of quoted parts (1131 × 1600 PNG): 20 datasheet pages (catalogue/datasheets/DS-01.png .. DS-20.png), 3 erratum pages (catalogue/errata/) and 4 lot arrival report pages (catalogue/lot_reports/); the catalogue's makers, item codes and values are fictional; the brief, the build log, the quote and the other documents, the hidden car and the grader stay in the benchmark repository; made for this task |
racecar_cad_design_from_photos |
ground | 87 photos of a tenth-scale four-wheel-drive autonomous race car (photos/photo_001.jpg .. photo_087.jpg, 2016 × 1512 JPEG; Blender Cycles renders of the benchmark maintainer's own model of the car in a lab room): its 23 made parts with paper tags on A3 and A2 cutting mats with scale references, both sides, and nine close-ups; the complete car on the bench from six views; 16 stage photos of its teardown; the rolling base assembly as one STEP file of 146 solids in the car's frame with its README.md (base/); for each of the 15 purchased component types (front wheel set, damper, differential, front drive shaft, foam bumper, 2D scanning LiDAR, compute module, motor controller, power distribution board, antenna and five spacer pins) under purchased/<part number>/ a STEP file of its geometry in its own frame, a datasheet (Markdown) and a drawing (2200 × 1300 PNG); parts_list.md, the made parts' tags, materials and densities; the STEP files hold the model's own solids, the documents are written for this task and the part numbers are invented; the true solids of the made parts and the grader stay in the benchmark repository |
racecar_full_assembly_tenth_scale_ur10e |
ground | payload/: the givens of the assembly task in which a UR10e-class arm with a two-finger gripper builds a tenth-scale race car from its rolling chassis and eight loose units: the six unit types' part files (given/parts/, STL, millimetres), five drawings (given/drawings/, PNG), the 14-photo picture manual with photos.json (given/manual/, renders with physical scale references), the arm and gripper MuJoCo models (given/robot/, BSD-3-Clause and BSD-2-Clause, their LICENSE files next to them) and render/, the agent cameras' scenery (the baked factory hall, the bench top's wood maps); given/robot/ and render/ are byte copies of the files of quadruped_full_assembly_12dof_ur10e, uploaded again under this task's prefix because each task's asset list names one prefix; verifier/gt.tar.xz: the grader's kit of the cell (kit.json, the units' meshes and collision pieces), downloaded by the verify stage, never in an image; made for this task from the benchmark maintainer's own model of the car |
racecar_time_trial_on_unseen_track |
ground | payload/: the givens of the behavior task in which a tenth-scale autonomous race car's program does a race day unattended at a venue it has never seen, staged into the agent home: the car's 14 laboratory driving logs (given/lab_logs/*.jsonl: every sensor, the commands and an overhead tracker's pose, recorded by the task's own lab rig), 12 photographs of two public example venues of the task's own venue generator (given/league/photos/, path-traced stills; never a hidden venue) and the car's outer geometry (given/car/meshes/: 13 OBJ parts, car.mtl, 3 textures and a README.md); the documents, the client library, the hidden venues and the grader stay in the benchmark repository; made for this task from the benchmark maintainer's own model of the car |
racecar_head_to_head_race_on_unseen_track |
ground | payload/: the givens of the behavior task in which a tenth-scale autonomous race car's program races a head-to-head day (a practice, then two heats of four cars from the back of the grid) unattended at a venue it has never seen, staged into the agent home: the car's 14 laboratory driving logs (given/lab_logs/*.jsonl), 16 photographs (given/league/photos/: 11 path-traced stills of two public example venues of the venue generator, never a hidden venue, and 5 path-traced scrutineering photographs of the car with the league's marker box on it) and the car's outer geometry (given/car/meshes/: 13 OBJ parts, car.mtl, 3 textures and a README.md); 43 of the 48 files are byte-identical copies of racecar_time_trial_on_unseen_track's (one photograph under a new name), the 5 scrutineering photographs are this task's own; the documents, the client library, the hidden venues and the grader stay in the benchmark repository; the car's meshes and the scrutineering photographs are made from the benchmark maintainer's own model of the car |
tensegrity_lot_acceptance_from_procurement_spec |
soft | 15 files that the agent reads while it turns the procurement specification of a cable-driven three-strut tensegrity scout into the acceptance tests of its production lots: 8 document page images (PNG): the site survey sheet (conops/ledge_survey.png), two design sheets (design/sheets/cable_map.png, stance_sheet.png) and the contract manufacturer's process flow, a traveler sample and three process-capability pages (manufacturing/); and the raw telemetry of six field runs with its channel list (field_reports/telemetry/RUN-A.csv .. RUN-F.csv, channels.md); the manufacturer, part numbers, lots and values are fictional, made for this task |
kitchen_appliance_tending |
arm | payload/render/: the kitchen around the cell as the agent cameras' renderer (MuJoCo's Filament) draws it; the image keeps it in a root-owned folder, it is not staged into the agent home. room.npz and room.json: the part of the apartment of the Unity Asset Store pack "ArchVizPRO Interior Vol.8 (URP)" (ArchVizPRO) that the task's cameras can see, as decimated meshes with base-colour maps of 256 to 512 px and one roughness and metallic value per material; ibl.ktx and sky.ktx: the image light and the sky behind the windows, from three OVRTX path-traced light probes of that kitchen, prefiltered with Filament's cmgen; made for this task. payload/units/: the three practice appliances that the agent may read and simulate (Stovetop018, Microwave075, Oven039: MJCF models with their visual meshes and textures, staged into the agent home), unmodified RoboCasa v1.0 fixture models (CC BY 4.0, from nvidia/PhysicalAI-Robotics-Manipulation-Objects-Kitchen-MJCF at 420a04af); attribution in the task's tests/NOTICE.md in the benchmark repository |
keyed_nut_peg_assembly |
arm | payload/render/: the laboratory around the cell as the agent cameras' renderer (MuJoCo's Filament) draws it; the image keeps it in a folder that only root and the cell's own service account read, it is not staged into the agent home. room.npz and room.json: the office scene of the Unity Asset Store pack "HQ ArchViz Loft Office Modular" and workshop furniture of "House Props Furniture Vol 2" (recoloured to grey and black), as decimated meshes with base-colour maps of 512 to 1024 px and one roughness and metallic value per material, with the cell's bench, pedestal and camera gantry; ibl.ktx: the image light, from an OVRTX path-traced light probe of that room over the bench (sky: Poly Haven's kloofendal_48d_partly_cloudy_puresky, CC0), prefiltered with Filament's cmgen; made for this task |
rov_requirements_engineering_from_tender |
underwater | 95 files that the agent reads while it evaluates eight bids for an inspection-class ROV at a reservoir intake: 71 page images (PNG): three drawings of the intake tower and the previous campaign's hand-written dive log (site/), five scanned factory test reports per bid (bids/T-01/test_reports/ to bids/T-08/test_reports/), 24 datasheets and 3 erratum sheets of the parts catalogue (catalogue/datasheets/, catalogue/errata/); 24 CSV logs (site/): the reservoir stage, the units' discharge, 18 CTD casts, the Secchi readings and the telemetry of three dives of the previous campaign; the site, the bidders, the makers, the part numbers and the values are fictional, made for this task |
rov_cad_design_from_photos |
underwater | 156 photos of an eight-thruster inspection-class ROV (photos/photo_001.jpg .. photo_156.jpg, 2016 × 1512 and 2400 × 1800 JPEG; path-traced renders of the vehicle in a workshop, made for this task): its 31 made parts with paper tags on cutting mats with scale references, both sides, 17 extra views and 35 close-ups; the complete vehicle on the bench from all around, from above and turned over; 32 stage photos of its teardown; for each of the 8 purchased module types (electronics enclosure, battery enclosure, a ducted thruster for each propeller hand, LED light, single-axis gripper, Doppler velocity log, ballast weight) a STEP file of the module's solids, a datasheet (Markdown) and a dimensioned drawing (PNG) under purchased/; the parts list parts_list.md; the vehicle in the photos and the STEP files derive from the maker's public CAD files of the vehicle, named in the task's tests/NOTICE.md in the benchmark repository |
rov_final_assembly_and_trim_single_arm |
underwater | payload/: the givens of the assembly task in which a UR10e-class arm with a wide parallel gripper ballasts an inspection-class ROV for the water named on its work order and fits its electronics enclosure, four vertical thrusters and two thruster guards: the six unit files (given/parts/, STL, millimetres), seven drawings (given/drawings/, PNG), the trim sheet, a water-density chart and a propeller data sheet (given/docs/, PNG), the 16-photo picture manual with photos.json (given/manual/, renders with physical scale references), the arm and gripper MuJoCo models (given/robot/, BSD-3-Clause and BSD-2-Clause, their LICENSE files next to them) and render/, the agent cameras' scenery (the baked factory hall, the bench top's wood maps); the arm model and render/ are byte copies of the files of quadruped_full_assembly_12dof_ur10e (but render/room/room.json, which lacks three top-level strings), uploaded again under this task's prefix because each task's asset list names one prefix; verifier/gt.tar.xz: the grader's kit of the cell (kit.json, the units' meshes and collision pieces), downloaded by the verify stage, never in an image; the unit files, the manual's renders and the kit's meshes derive from the maker's public CAD files of the vehicle, named in the task's tests/NOTICE.md in the benchmark repository |
longtail_manipulation_recovery |
arm | two archives of unmodified RoboCasa v1.0 assets (CC BY 4.0, Nasiriany et al., RoboCasa, RSS 2024: the fixtures and Lightwheel objects of nvidia/PhysicalAI-Robotics-Manipulation-Objects-Kitchen-MJCF at 420a04af, the Objaverse objects and textures of robocasa/robocasa-assets), unpacked by the image builds into the simulator's asset tree and not staged into the agent home: payload/robocasa_assets.tar.gz: the 1,971 files that the practice kitchens of the agent's image load (fixtures, textures, 55 object instances, RoboCasa's training layouts and styles); payload/verifier_assets.tar.gz: the 2,309 files that the graded kitchens load (RoboCasa's held-out layouts and styles, 82 other object instances), fetched by the task's separate verifier image only and never by the agent's; these held-out files are public RoboCasa files and are here by the owner's exception (2026-11-17) to the rule that the dataset holds agent-visible files; payload/look_assets.tar.gz: the two files of the cameras' colour renderer, fetched by both images: ibl.ktx, an image light of a neutral room generated for this task (no third-party content), and city_card.png, the picture behind the kitchens' window panes, a crop of Poly Haven's sunset_jhbcentral panorama (CC0); attribution in the task's tests/NOTICE.md in the benchmark repository |
tensegrity_requirements_replica_from_photos |
soft | 320 photos of one strut of a three-strut, cable-driven tensegrity rolling robot on an electronics bench, complete and in eight teardown stages down to its made parts with paper tags on a cutting mat (2016 × 1512 JPEG, Blender Cycles renders of the designers' published whole-strut mesh in a lab room; scale references: cutting mat, checker cards, steel rules). The robot is the open-source three-strut, cable-driven tensegrity robot of the Yale Faboratory (Rebecca Kramer-Bottiglio, Yale University) and the PRACSYS lab (Kostas Bekris, Rutgers University): W. R. Johnson III, P. Meng, N. Chen, L. Cimatti, A. Vercoutere, M. Aanjaneya, R. Kramer-Bottiglio, K. E. Bekris, "An Open-Source, Reproducible Tensegrity Robot that can Navigate Among Obstacles", IEEE Robotics and Automation Letters 11(6):7054-7061, 2026 (arXiv:2511.05798); the strain sensors and their cut outlines come from W. R. Johnson III et al., "Impact-resistant, autonomous robots inspired by tensegrity architecture" (arXiv:2501.15078) and its figshare dataset. Source files, licences and what the task does with them: the task's tests/NOTICE.md in the benchmark repository |
tensegrity_cad_design_from_photos |
soft | 287 photos of one strut of a three-strut, cable-driven tensegrity rolling robot on a lab bench: the assembled strut, eight stages of taking it apart and every made part laid out with paper tags on a cutting mat (2016 × 1512 JPEG, Blender Cycles renders of the designers' published whole-strut mesh in a lab room, byte for byte 287 of the photos of tensegrity_requirements_replica_from_photos; scale references: cutting mat, checker cards, a steel rule), and the strut's two purchased components under purchased/, each with its STEP file, a data sheet page (Markdown) and a dimensioned drawing (PNG) made for the task: the micro metal gearmotor (Pololu item 5192; the vendor's STEP model written again after a rigid move) and the 1/8 in × 12 in aluminium rod (McMaster-Carr item 8974K19; a faceted solid converted triangle for triangle from the published mesh). The robot, its designers and their papers are named in the row of tensegrity_requirements_replica_from_photos. Source files, licences and what the task does with them: the task's tests/NOTICE.md in the benchmark repository |
tensegrity_strut_assembly_3bar_panda |
soft | payload/: the givens of the assembly task in which a Franka Panda finishes the three struts of a small three-strut, cable-driven tensegrity rolling robot from fifteen loose modules: the five module files (given/parts/, STL, millimetres; the robot designers' published whole-strut mesh grouped into modules, triangle for triangle, each file in a frame of its own), the 25-photo picture manual with photos.json (given/manual/, 2016 × 1512 JPEG, Blender Cycles renders of the same triangles with physical scale references in view), the Franka Panda MuJoCo model (given/panda/, Apache-2.0, from MuJoCo Menagerie, its LICENSE included; the same files as robotic_arm_full_assembly_so101_panda's) and render/, the scenery of the agent cameras' renderer (room/: a university laboratory built from the Archviz Laboratory Pack, Unity Asset Store, with its Blender Cycles lightmaps and image light; bench/: the bench top's laminate maps, generated by code); verifier/gt.tar.xz: the grader's kit of the cell (kit.json, the mating graph, the modules' meshes and carved collision pieces), downloaded by the verify stage, never in an image. The robot, its designers and their papers are named in the row of tensegrity_requirements_replica_from_photos. Source files, licences and what the task does with them: the task's tests/NOTICE.md in the benchmark repository |
lab_sample_handling |
arm | payload/render/: the laboratory around the bench as the bench cameras' renderer (MuJoCo's Filament) draws it; the images keep it in a folder that only root and the commissioning service's own account read, it is not staged into the agent home. room.npz and room.json: a bench bay of the Unity Asset Store pack "Archviz Laboratory Pack" (wall benches, shelf structures, bench-top glassware and boxes, ceiling luminaires; the tube racks left out), as decimated meshes (226,397 triangles) with base-colour maps and one roughness and metallic value per material, with the task bench's own worktop and cabinet; ibl.ktx: the image light, from an OVRTX path-traced light probe of that bay over the bench (sky: Poly Haven's kloofendal_48d_partly_cloudy_puresky, CC0) with the bay's ceiling luminaires painted in, prefiltered with Filament's cmgen; made for this task |
freeflyer_cad_design_from_photos |
space | 119 photos of a cube-shaped cabin free-flyer (a core module between two propulsion modules with corner-bumper sets, a perching arm with a two-finger gripper, stowed) on a lab bench: the assembled unit, three teardown stages, every part on a cutting mat from both sides with a checker card and a steel rule, close-ups, two more arm poses and the gripper open (photos/photo_001.jpg .. photo_119.jpg, 2016 × 1512 and 2400 × 1800 JPEG; Blender Cycles renders of the robot makers' published appearance meshes in a lab room, made for this task; the sources and their licences are in the task's tests/NOTICE.md in the benchmark repository) |
freeflyer_uncrewed_station_caretaking |
space | payload/: the givens' pictures and the simulation cell's models of the behavior task in which a cabin free-flyer's program works alone in an uncrewed three-module station at orbital night; the image keeps them in a root-owned folder that the agent's account cannot read. given/: 40 pictures that the task's setup stages into the agent home: 6 figures of the robot manual and the station handbook (given/figures/, PNG), 16 photographs of the lost-item catalogue (given/catalogue/, JPEG) and 18 views of the as-surveyed station map (given/station_map/views/, PNG: a plan, an elevation and 16 wall views); cell/: the three archives of the cell's models, unpacked by the image build into the asset root that the rehearsal facility and the grader read: robot_model_v1_2.tar.xz (the robot's collision model with a light visual set, and the dock: MJCF and meshes reconstructed for this task from the makers' public documents), station_look_v05.tar.xz (the display meshes and textures of the station modules and the dock, made for this task from public 3D models of a station interior with signage, part numbers and insignia removed) and robot_author_v1_1.tar.xz (the robot's visible meshes as its makers published them with the robot's open-source flight software, Apache License 2.0, converted in format, units and frame only); the documents, the client library, the hidden sites and the grader stay in the benchmark repository; the sources and their licences are in the task's tests/NOTICE.md in the benchmark repository |
pantry_jar_refilling |
arm | payload/render/: the kitchen around the counter as the cameras' renderer (MuJoCo's Filament) draws it; the image keeps it in /opt/task-assets, readable by every account (the agent's own copy of the simulator draws with it), it is not staged into the agent home. room.npz and room.json: a living room and kitchen of the Unity Asset Store pack "ArchVizPRO Interior Vol.6" (the house's shell, the kitchen run with its hood, sink, fridge and what stands on it, the dining area), as decimated meshes (266,812 triangles) with base-colour maps and one roughness and metallic value per material, with the task's counter as the kitchen's island (a top of oak staves composed from Poly Haven's oak_veneer_05, CC0; lacquered fronts; a plinth); ibl.ktx: the image light, from an OVRTX path-traced light probe of that kitchen over the counter (sky: Poly Haven's kloofendal_48d_partly_cloudy_puresky, CC0) with the room's lights painted in, prefiltered with Filament's cmgen; made for this task |
Licence
CC BY 4.0, except files that carry their own licence (the UR10e and 2F-85 models of quadruped_full_assembly_12dof_ur10e and of racecar_full_assembly_tenth_scale_ur10e (payload/given/robot/, the same files): BSD-3-Clause and BSD-2-Clause, their LICENSE files next to them; its actuator STEP files: modified derivatives of a BSD-3-Clause design, its notice in that task's tests/NOTICE.md in the benchmark repository; the Franka Panda model of robotic_arm_full_assembly_so101_panda and robotic_arm_full_assembly_openarm_panda: Apache-2.0, its LICENSE file next to it; the UFACTORY xArm 7 model of uav_assembly_agipix_xarm7: BSD-3-Clause, from MuJoCo Menagerie, its LICENSE file next to it; the OpenArm units of robotic_arm_full_assembly_openarm_panda: modified derivatives of the CERN-OHL-S-2.0 OpenArm 2.0 design (Enactic), distributed under that licence, its notice and the statement of modifications in that task's tests/NOTICE.md in the benchmark repository; the actuator and foot STEP files of quadruped_cad_design_from_photos: modified derivatives of a BSD-3-Clause design, its notice and licence text in that task's NOTICE.md; the bought parts' STEP files of bimanual_cad_design_from_photos: modified derivatives of a CERN-OHL-S-2.0 open-hardware design, distributed under that licence, its notice and the statement of modifications in that task's verify/NOTICE.md in the benchmark repository; the photo set of bimanual_requirements_replica_from_photos: renders of a modified derivative of a CERN-OHL-S-2.0 open-hardware design, distributed under that licence, its notice, the date and the statement of modifications in that task's tests/NOTICE.md in the benchmark repository; the published parts' STL files of robotic_arm_robot_description_from_photos (parts/P01.stl-P10.stl, parts/base_in_world.stl): modified derivatives of an Apache-2.0 open-hardware design, its notice and the statement of modifications in that task's tests/NOTICE.md in the benchmark repository; the clips of hot3d_sharpa_retarget: adapted from the HOT3D dataset's sequence data, CC BY-SA 4.0, tasks/hand/hot3d_sharpa_retarget/LICENSE-HOT3D.txt; attribution: HOT3D, Meta Platforms Technologies, LLC, Banerjee et al., arXiv:2406.09598). The renders of robocasa_object_dynamics_from_video show RoboCasa's CC BY 4.0 assets; its attribution is in that task's verify/NOTICE.md. The practice units of kitchen_appliance_tending (payload/units/) are RoboCasa's CC BY 4.0 fixture models; their attribution is in that task's tests/NOTICE.md in the benchmark repository. The SuperDex assets of dexhand_shape_sorter_all_pieces: the shape_box prefab CC BY 4.0 and the Wuji Hand 2 models MIT (Wuji Technology, its NOTICE), their LICENSE files next to them. The vendors, part numbers and datasheet values in the document images are fictional. The files of driving_safety_supervisor and aerial_search_geolocation are byte-for-byte mirrors of third-party releases and keep the terms of their sources (CARLA: MIT code and CC-BY assets, Unreal Engine 4 under Epic's terms; the other sources as named in the task table). The files of novel_object_6dof_grasp_detection are a derived subset of GraspNet-1Billion (Fang, Wang, Gou and Lu, CVPR 2020, https://graspnet.net) and keep its non-commercial terms (CC BY-NC-SA 4.0), stated in the NOTICE next to them. The photo of robotic_arm_environment_twin_from_real_photo is from the HOPE dataset (NVIDIA, CC BY-NC-SA 4.0); its attribution is in that task's tests/NOTICE.md in the benchmark repository. The photos of quadruped_environment_twin_from_real_photos are from the ETH3D benchmark (Schöps, Schönberger, Galliani, Sattler, Schindler, Pollefeys and Geiger, CVPR 2017, https://www.eth3d.net/) and keep its licence (CC BY-NC-SA 4.0); their attribution is in that task's tests/NOTICE.md in the benchmark repository. The part files, picture manual and drawings of racecar_full_assembly_tenth_scale_ur10e and the meshes of its verifier/gt.tar.xz are made from the same model of the maintainer's (the photos rendered from it, the meshes decimated from it, two product-label textures replaced by unlabelled copies); what derives from the makers' CAD files keeps their terms, as that task's tests/NOTICE.md in the benchmark repository states. The lab logs, photographs and car meshes of racecar_time_trial_on_unseen_track are made for this task (the logs recorded by its own lab rig, the photographs rendered for it, the meshes exported from the same model of the maintainer's car with neutral part names); what derives from the makers' CAD files keeps their terms, as that task's tests/NOTICE.md in the benchmark repository states. The arm and gripper models of rov_final_assembly_and_trim_single_arm (payload/given/robot/) are BSD-3-Clause (the UR10e model, the same files as quadruped_full_assembly_12dof_ur10e's) and BSD-2-Clause (its wide gripper, derived from a BSD-2-Clause MJCF model), their LICENSE files next to them. Its unit files (payload/given/parts/), the renders of its picture manual and the meshes of its verifier/gt.tar.xz derive from the maker's public CAD files of the vehicle (named in that task's tests/NOTICE.md in the benchmark repository); what derives from the maker's CAD files keeps their terms. Its drawings, trim sheet, density chart and propeller data sheet are made for this task (the density chart computed with the TEOS-10 GSW routines, BSD-style licence, its notice in the same NOTICE.md). The STEP files of rov_cad_design_from_photos (payload/purchased/) and the vehicle and parts in its photos derive from the maker's public CAD files of the vehicle (named in that task's tests/NOTICE.md in the benchmark repository); what derives from the maker's CAD files keeps their terms. Its datasheets, drawings and parts list are made for this task. The photos of tensegrity_requirements_replica_from_photos and tensegrity_cad_design_from_photos are renders of the robot designers' published whole-strut mesh, taken from a repository that has no licence file (the package.xml of its branch says MIT), and the rod STEP file of tensegrity_cad_design_from_photos is one piece of that mesh; its gearmotor STEP file is the vendor's model written again after a rigid move and keeps the vendor's terms. The sources, what each states about its licence and the attribution are in each task's tests/NOTICE.md in the benchmark repository. The module files of tensegrity_strut_assembly_3bar_panda (payload/given/parts/), the renders of its picture manual and the meshes of its verifier/gt.tar.xz are the same published whole-strut mesh, grouped into modules (the repository it was taken from has no licence file; the package.xml of its branch says MIT); its Franka Panda model (payload/given/panda/, the same files as robotic_arm_full_assembly_so101_panda's) is Apache-2.0, its LICENSE file next to it. The sources, what each states about its licence and the attribution are in that task's tests/NOTICE.md in the benchmark repository. The photographs of racecar_requirements_replica_from_photos are renders of the benchmark maintainer's own model of the car, which he supplied for public distribution, in a lab room scene supplied on the same terms; its catalogue pages are made for this task and describe no real product, as that task's tests/NOTICE.md in the benchmark repository states. The photographs, STEP files and documents of racecar_cad_design_from_photos are made for this task from the benchmark maintainer's own model of the car, which he supplied for the benchmark and agreed to distribute publicly in these forms: the photographs are renders of it, the STEP files hold its own solids (with the lettering on five of them removed) and the documents are written from its measured geometry, as that task's tests/NOTICE.md in the benchmark repository states. The lab logs, photographs and car meshes of racecar_head_to_head_race_on_unseen_track are, for 43 of its 48 files, unchanged copies of racecar_time_trial_on_unseen_track's and carry the same terms; its five scrutineering photographs are renders made for this task from the same model of the maintainer's car; what derives from the makers' CAD files keeps their terms, as each task's tests/NOTICE.md in the benchmark repository states. The models under cross_embodiment_adaptation's payload/visual_assets/ keep their sources' licences: MuJoCo Menagerie's Go1 (BSD-3-Clause) and Crazyflie 2 (MIT), their LICENSE files next to them, and Poly Haven's textures and models (CC0-1.0); provenance.json lists each file's source.
- Downloads last month
- 2,625