Sandbox crafting & building
Gather, craft and build structures to a blueprint.
20 project-authored proposals · taxonomy v1.1 · 0 new benchmark runs. Verifiers and partial-credit rules below are designs, not executed results.
- Typical applications
- Minecraft · Teardown
- Class GUI-need label
G3· GUI-only
Task proposals · 20
Expand to inspect the complete original wording and proposed checks. Original proposal text remains English in both interfaces.
R13.01.02.03-T01 Craft a wooden pickaxe from scratch in a fresh Minecraft world
Original instruction
A Minecraft Java 1.20.4 survival world (seed 123, peaceful) is loaded at spawn with an empty inventory. Punch trees, craft planks, sticks and a crafting table, and craft a wooden pickaxe.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Minecraft
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clickdrag· Drag & dropvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Server plugin reads player inventory: wooden_pickaxe present; crafting statistics show it was crafted (not given). Commands and creative mode disabled.
Proposed partial credit
20% logs collected; 20% planks and sticks; 20% crafting table placed; 40% pickaxe crafted
R13.01.02.03-T02 Mine and smelt three iron ingots in Minecraft
Original instruction
Minecraft Java survival (seed 123, peaceful), player starts with a stone pickaxe, 16 coal and 32 cobblestone. Find and mine 3 iron ore, craft a furnace and smelt them into 3 iron ingots.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Minecraft
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
click· Point & clickdrag· Drag & dropnav· Pan/zoom/scroll- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Server stats: mined iron_ore/deepslate_iron_ore ≥ 3; inventory iron_ingot ≥ 3; furnace crafted and used (stat 'interact_with_furnace' ≥ 1). Commands disabled.
Proposed partial credit
40% scaled by ore mined; 20% furnace crafted; 40% scaled by ingots
R13.01.02.03-T03 Craft a bed with the Minecraft recipe book
Original instruction
Minecraft Java survival, standing at a crafting table with 3 white wool and 3 oak planks in the inventory. Open the crafting table, use the recipe book to fill the bed recipe and craft one white bed.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Minecraft
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
click· Point & clickvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 1 / 5
Proposed verifier
Inventory contains white_bed; crafting statistics increment for white_bed. Commands, cheats and creative mode are disabled for the agent on the server, so the bed cannot be given by command and must be crafted through the crafting GUI.
Proposed partial credit
40% crafting GUI opened with ingredients placed; 60% bed crafted
R13.01.02.03-T04 Build a 5×5×4 cobblestone house at marked coordinates
Original instruction
Minecraft Java survival world with 200 cobblestone, 40 oak planks, an oak door and glass. Corner markers (gold blocks) show a 5×5 footprint. Build walls 4 high in cobblestone with a door on the south side, 2 glass windows, and a flat plank roof.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Minecraft
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
geom· Precise geometryclick· Point & clicknav· Pan/zoom/scroll- Work product
WP04· 3D/spatial model- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Server exports the build region as a structure; voxel comparison to the template: IoU ≥ 0.9 on block type classes (cobblestone walls, plank roof, door south, ≥ 2 glass). Commands and WorldEdit unavailable.
Proposed partial credit
40% scaled by wall IoU; 20% roof; 20% door on south; 20% windows
R13.01.02.03-T05 Replicate a structure from a reference screenshot
Original instruction
Minecraft Java survival (resources in chests) beside an empty plot. input/target.png shows a small two-storey tower with stairs, slabs and a balcony from one angle; the back is not visible but is symmetric. Build it on the plot to match.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Minecraft
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
vis· Visual judgmentgeom· Precise geometrynav· Pan/zoom/scroll- Work product
WP04· 3D/spatial model- Conditions
long· Long-horizon- Proposed difficulty
- 5 / 5
Proposed verifier
Build region exported and compared to the hidden template: voxel IoU (block type class aware) ≥ 0.85 for full credit, with partial credit by IoU. Commands, WorldEdit and structure blocks are disabled.
Proposed partial credit
50% scaled by voxel IoU from 0.4 to 0.625; 50% scaled from 0.625 to 0.85
R13.01.02.03-T06 Build a walkable bridge across an eight-block river
Original instruction
Minecraft Java survival, player on the west bank of an 8-block-wide river with 64 oak planks and 32 fences. Build a bridge with railings that lets a player walk across without swimming or jumping gaps.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Minecraft
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
geom· Precise geometryclick· Point & clicknav· Pan/zoom/scroll- Work product
WP04· 3D/spatial model- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Path-finding check on the exported region: a walkable path exists from west to east bank above water level with no 2-block gaps; fences on both sides along ≥ 80% of the span. Commands, WorldEdit and creative mode are disabled for the agent on the server, so blocks cannot be placed by command.
Proposed partial credit
60% walkable crossing; 40% railings coverage
R13.01.02.03-T07 Light a room so no spot is dark enough for mobs
Original instruction
Minecraft Java survival, an enclosed 12×12 stone room with 8 torches in inventory. Place torches so the block light level is at least 8 on every floor block (mobs cannot spawn).
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Minecraft
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
geom· Precise geometryclick· Point & clickvis· Visual judgment- Work product
WP04· 3D/spatial model- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Server light-level scan of all 144 floor blocks: minimum block light ≥ 8; torches placed ≤ 8. F3 light readout is disabled, so placement needs spacing reasoning or visual brightness cues.
Proposed partial credit
50% scaled by fraction of floor blocks with light ≥ 8 from 0 to 0.5; 50% scaled from 0.5 to 1.0
R13.01.02.03-T08 Demolish only a marked wall section in Teardown
Original instruction
Teardown sandbox map with a brick building; a painted red rectangle marks a 3×2 m wall section to remove for a doorway. Using the sledgehammer and blowtorch, remove the marked section while removing less than 5% voxel volume outside it and keeping the building standing.
- Platforms
win· Windows- Applications
- Teardown
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
geom· Precise geometryrt· Real-time controlvis· Visual judgment- Work product
WP04· 3D/spatial model- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Harness mod exports the wall voxel grid after the run: ≥ 95% of marked voxels removed; removed voxels outside the mark ≤ 5% of the mark's volume; no structural collapse event. Only the harness mod is loaded (no other Lua mods or console for the agent), and the agent has no shell access to the save folder.
Proposed partial credit
50% scaled by marked voxels removed; 30% collateral limit; 20% building intact
R13.01.02.03-T09 Build valid NPC housing in Terraria
Original instruction
Terraria world (journey mode off, fixed seed) with wood, a workbench, torches and a door in the inventory. Build a house that passes the housing check (enclosed, background walls, light source, table, chair, door) and have the Guide move in.
- Platforms
win· Windowsmac· macOSlin· Linux- Applications
- Terraria
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
click· Point & clickgeom· Precise geometrydrag· Drag & drop- Work product
WP04· 3D/spatial model- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
World file parsed (TEdit-compatible reader): a valid house region by the game's rules and the Guide's home tile inside it. TEdit and tModLoader are not installed, the agent has no shell, and world-file writes are logged by process, so a world not saved by Terraria is detected.
Proposed partial credit
40% enclosure and walls; 30% furniture and light; 30% Guide assigned
R13.01.02.03-T10 Build a comfortable shelter in Valheim
Original instruction
Valheim local world, player has wood, stone, a hammer and resin. Build a shelter with a workbench, a fire under a roof with a vent so smoke escapes, a bed, and walls/roof so the 'Sheltered' status is active. Reach comfort level ≥ 3 while resting.
- Platforms
win· Windowslin· Linux- Applications
- Valheim
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
geom· Precise geometryclick· Point & clicknav· Pan/zoom/scroll- Work product
WP04· 3D/spatial model- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
World save parsed for pieces (workbench, fire, bed, roof); harness mod logs player status effects: Sheltered and Rested with comfort ≥ 3 while inside, and no 'Smoked' status. Devcommands are disabled (BepInEx not loaded for the agent) and the agent has no shell access to the world folder.
Proposed partial credit
25% structure pieces; 25% Sheltered; 25% comfort ≥ 3; 25% no smoke damage
R13.01.02.03-T11 Build a working 2×2 piston door with redstone
Original instruction
Minecraft Java survival with a chest of sticky pistons, redstone dust, repeaters, torches, stone and a lever. In a stone wall, build a flush 2×2 piston door that opens when the lever is switched on and closes when it is off.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Minecraft
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
geom· Precise geometryclick· Point & clickvis· Visual judgment- Work product
WP04· 3D/spatial model- Conditions
- None declared
- Proposed difficulty
- 5 / 5
Proposed verifier
Harness toggles the lever via a server test and checks the 4 doorway blocks: air when on, solid and flush with the wall when off, over 3 cycles. Commands disabled for the agent.
Proposed partial credit
30% pistons placed around doorway; 40% opens on; 30% closes flush on off
R13.01.02.03-T12 Set up a level-30 enchanting table
Original instruction
Minecraft Java survival with an enchanting table, 15 bookshelves, 30 levels of XP and lapis. Place the bookshelves so the enchanting table offers a level-30 enchantment, then enchant the provided diamond pickaxe with the level-30 option.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Minecraft
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
geom· Precise geometryclick· Point & click- Work product
WP04· 3D/spatial model- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Region export: 15 bookshelves at valid positions (one block gap, same or one level higher); item NBT shows enchantments from a level-30 roll (XP decreased by 3 levels from 30). Commands, cheats and creative mode are disabled for the agent on the server, so XP, items and blocks cannot be set by command.
Proposed partial credit
50% valid bookshelf arrangement; 50% level-30 enchant applied
R13.01.02.03-T13 Build and light a Nether portal using the bucket method
Original instruction
Minecraft Java survival next to a lava pool, with an iron bucket, a water bucket, flint and steel, and cobblestone. Create 10 obsidian by pouring water on lava source blocks, build a portal frame and light it.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Minecraft
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
geom· Precise geometryclick· Point & clickvis· Visual judgment- Work product
WP04· 3D/spatial model- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Server: an active nether_portal block cluster exists in the region; obsidian obtained by placement, not pickup (stats). Player deaths = 0 (lava risk). Commands, cheats and creative mode are disabled for the agent on the server, so obsidian and fire cannot be given or placed by command.
Proposed partial credit
40% obsidian created; 30% frame built; 30% portal lit without dying
R13.01.02.03-T14 Build a hydrated 9×9 wheat farm with fences
Original instruction
Minecraft Java survival with a hoe, water bucket, 80 wheat seeds and 40 fences. Build a 9×9 farm with a single water source in the centre so every farmland block is hydrated, plant wheat on all 80 farmland blocks and fence the perimeter with a gate.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Minecraft
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
geom· Precise geometryclick· Point & click- Work product
WP04· 3D/spatial model- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Region export: 80 farmland blocks with moisture 7 and wheat crops; water at centre; closed fence perimeter with a gate (flood-fill check). Commands, cheats and creative mode are disabled for the agent on the server, so blocks cannot be placed by command.
Proposed partial credit
30% farmland hydrated; 30% wheat planted; 40% enclosed fence with gate
R13.01.02.03-T15 Build a boss arena with platforms and buffs in Terraria
Original instruction
Terraria world with wood, platforms, a campfire and a heart lantern in inventory. Build an arena at least 100 tiles wide with 3 rows of platforms spaced 10-14 tiles apart vertically, and place the campfire and heart lantern so both buffs are active at the arena centre.
- Platforms
win· Windowsmac· macOSlin· Linux- Applications
- Terraria
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
drag· Drag & dropgeom· Precise geometrynav· Pan/zoom/scroll- Work product
WP04· 3D/spatial model- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
World file parse: platform rows width ≥ 100 tiles, 3 rows with spacing 10-14; campfire and heart lantern within buff range of the centre (buff check with in-game radii). TEdit and tModLoader are not installed, the agent has no shell, and world-file writes are logged by process, so a world not saved by Terraria is detected.
Proposed partial credit
40% platform rows; 20% spacing; 40% buffs reach centre
R13.01.02.03-T16 Recreate a 16×16 pixel-art image in coloured wool
Original instruction
Minecraft Java survival with chests of all 16 wool colours. input/pixel_heart.png is a 16×16 image. Build it as a vertical wall of wool facing south starting at the gold block marker, one block per pixel, choosing the closest wool colour per pixel.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Minecraft
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
geom· Precise geometryvis· Visual judgmentclick· Point & click- Work product
WP04· 3D/spatial model- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Region export compared per pixel to the expected wool colour (nearest by ΔE2000 from the palette): accuracy ≥ 98% for full credit; position must align with the marker. Commands, WorldEdit and creative mode are disabled for the agent on the server, so wool cannot be placed by command.
Proposed partial credit
20% correct position and orientation; 80% scaled by per-pixel accuracy
R13.01.02.03-T17 Build a vehicle in Besiege that carries a payload 100 metres
Original instruction
Besiege sandbox level with a 2-block payload crate. Build a machine with a starting block, wheels, a frame and a holder so it drives the crate past the 100 m line on uneven ground without dropping it.
- Platforms
win· Windowsmac· macOSlin· Linux- Applications
- Besiege
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
geom· Precise geometrydrag· Drag & droprt· Real-time control- Work product
WP04· 3D/spatial model- Conditions
- None declared
- Proposed difficulty
- 5 / 5
Proposed verifier
Harness mod logs crate and machine positions: crate x ≥ 100 m and above ground on the machine at the end. Machine file (.bsg) saved for inspection. No mods besides the harness logger are loaded and the agent has no shell, so positions cannot be set outside the simulation.
Proposed partial credit
30% machine moves under control; 70% scaled by crate distance up to 100 m
R13.01.02.03-T18 Survive the first Minecraft night with a shelter and a bed
Original instruction
Minecraft Java survival (normal), fresh spawn at dawn, empty inventory. Before nightfall, gather wood and wool, build an enclosed shelter, craft a bed, and sleep through the first night inside it.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Minecraft
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
click· Point & clickdrag· Drag & dropnav· Pan/zoom/scroll- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Server: player slept (stat sleep_in_bed ≥ 1) during night 1; bed inside an enclosed region (flood-fill shows no opening to the outside); deaths = 0. Commands, cheats and creative mode are disabled for the agent on the server, so items, blocks and time cannot be set by command.
Proposed partial credit
30% bed crafted; 30% enclosed shelter; 40% slept through night without dying
R13.01.02.03-T19 Build a tower from a layer-by-layer schematic PDF
Original instruction
Minecraft Java survival with resources in chests. input/tower_layers.pdf shows 5 layers of a 7×7 tower as coloured grids with a block legend. Build it at the marker so each layer matches, including windows and the stair entrance.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- Minecraft · PDF viewer
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
geom· Precise geometrymulti· Cross-appnav· Pan/zoom/scroll- Work product
WP04· 3D/spatial model- Conditions
multiapp· Multi-app- Proposed difficulty
- 4 / 5
Proposed verifier
Region export vs schematic: per-layer block-type accuracy; full credit at ≥ 95% overall with stair orientation correct. Commands, WorldEdit and creative mode are disabled for the agent on the server, so blocks cannot be placed by command.
Proposed partial credit
80% scaled by block accuracy (one fifth per layer); 20% stair orientation
R13.01.02.03-T20 Craft and fill a chest in Minecraft Bedrock on Android
Original instruction
On the Android emulator, Minecraft Bedrock (survival world) is open with 8 oak planks, 5 apples and 3 torches. Using the touch controls, craft a chest, place it on the marked grass block, and put the apples and torches into it.
- Platforms
and· Android- Applications
- Minecraft
- Interfaces
mob· Mobile touch- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
touch· Touch gesturesdrag· Drag & drop- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
World pulled from the emulator and parsed (LevelDB): chest at the marked coordinates containing 5 apples and 3 torches. Cheats are off in the world settings and adb/root are not exposed to the agent, so the world cannot be edited directly; input comes from emulator touches.
Proposed partial credit
40% chest crafted; 30% placed on marker; 30% items stored
Mapped research references
Existing benchmark-to-class links from the research catalog. They are not record-level annotations, executions or product endorsements.