Racing & vehicle control
Steer vehicles through courses in games and sims.
20 project-authored proposals · taxonomy v1.1 · 0 new benchmark runs. Verifiers and partial-credit rules below are designs, not executed results.
- Typical applications
- racing games · BeamNG
- 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.
R13.01.01.03-T01 Finish a SuperTuxKart race on the podium
Original instruction
SuperTuxKart is set to a single race on 'Lighthouse', 3 laps, Novice difficulty, 4 AI karts, your kart 'Tux'. Finish in the top 3 using steering, nitro and items with the keyboard.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- SuperTuxKart
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
rt· Real-time controlkey· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Race result written by a harness build of SuperTuxKart (results.xml): finishing position ≤ 3 and 3 laps completed. The agent sends OS key events only; game console and artist debug mode are disabled.
Proposed partial credit
40% race finished; 60% scaled by position (5th 0, 3rd full)
R13.01.01.03-T02 Lap a TORCS speedway under 60 seconds without leaving the track
Original instruction
TORCS runs a practice session on 'CG Speedway number 1' in the car 'car1-trb1' with automatic gearbox. Complete one flying lap under 60.0 s without any wheel leaving the asphalt.
- Platforms
lin· Linuxwin· Windows- Applications
- TORCS
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
rt· Real-time controlkey· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Telemetry from the TORCS harness robot interface (read-only): best valid lap time < 60.0 s with no off-track samples during that lap. Agent controls via keyboard only.
Proposed partial credit
30% clean lap completed; 70% scaled by lap time from 75 s to 60 s
R13.01.01.03-T03 Park a car in a marked bay in BeamNG.drive
Original instruction
BeamNG.drive scenario: a sedan on the West Coast USA parking lot must be reverse-parked into the marked bay between two parked cars. End stopped with the car centred within 0.3 m and heading within 5° of the bay axis, without touching anything.
- Platforms
win· Windows- Applications
- BeamNG
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
rt· Real-time controlgeom· Precise geometryvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
BeamNGpy (harness side only) reads final vehicle pose and damage: lateral and longitudinal offset ≤ 0.3 m, heading error ≤ 5°, damage = 0. The Lua console and world editor are disabled for the agent.
Proposed partial credit
20% stopped inside the bay; 40% scaled by position error; 20% heading; 20% no contact
R13.01.01.03-T04 Beat a target time at Hirochi Raceway in BeamNG
Original instruction
BeamNG.drive time trial on Hirochi Raceway (full circuit) in the provided sports car, traction control on. Set a valid lap under 1:58.0 within 6 attempts; cutting chicanes invalidates a lap.
- Platforms
win· Windows- Applications
- BeamNG
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
rt· Real-time controlkey· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Scenario lap timer and track-limits checker (harness Lua extension) log valid lap times; best valid lap compared to 1:58.0. The Lua console and world editor are disabled for the agent and BeamNGpy runs harness-side only, so vehicle state and timers cannot be set by script.
Proposed partial credit
20% valid lap; 80% scaled by best valid lap from 2:15 to 1:58
R13.01.01.03-T05 Finish a lap under three minutes in The Need for Speed (DOS)
Original instruction
DOSBox runs The Need for Speed (harness-supplied), practice mode on 'Rusty Springs' in the Mazda RX-7. Complete one lap under 3:00.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- racing games
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
rt· Real-time controlkey· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Lap counter and timer read from DOS memory by the harness; fallback OCR of the HUD timer. Pass at lap time < 3:00. The DOSBox debugger and config file are not accessible to the agent, so DOS memory cannot be patched; input is OS key events only.
Proposed partial credit
40% lap completed; 60% scaled by lap time from 4:00 to 3:00
R13.01.01.03-T06 Earn a bronze medal on the first Trackmania Nations Forever track
Original instruction
Trackmania Nations Forever is on campaign track A01 (offline). Drive a full run and earn at least the bronze medal time. Retries are allowed.
- Platforms
win· Windows- Applications
- Trackmania
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
rt· Real-time controlkey· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 2 / 5
Proposed verifier
Replay file (.Replay.Gbx) saved by the game is parsed: finish time ≤ bronze time and checkpoint count complete. Replays are validated by the game's own replay validation.
Proposed partial credit
40% finish; 60% bronze or better
R13.01.01.03-T07 Win a four-race SuperTuxKart grand prix on Intermediate
Original instruction
SuperTuxKart grand prix 'Penguin Playground' (4 tracks), Intermediate difficulty, 5 AI karts. Finish the cup in first place on total points.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- SuperTuxKart
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
rt· Real-time controlkey· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
long· Long-horizon- Proposed difficulty
- 5 / 5
Proposed verifier
Grand-prix result file: player rank 1 on total points after 4 races. Debug features disabled.
Proposed partial credit
60% races finished top 3 (a quarter per race); 40% overall cup win
R13.01.01.03-T08 Finish a 10-lap Speed Dreams race with a planned pit stop
Original instruction
Speed Dreams quick race on 'Espie' in a car-ls1 with damage and fuel consumption on, starting with fuel for 6 laps, 10 laps total, 5 AI cars. Make a pit stop with a fuel amount set in the pit menu and finish the race without running dry, in the top 4.
- Platforms
lin· Linuxwin· Windowsmac· macOS- Applications
- Speed Dreams
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
rt· Real-time controlkey· Shortcutsclick· Point & click- Work product
WP11· Game/sim state- Conditions
long· Long-horizon- Proposed difficulty
- 5 / 5
Proposed verifier
Race result XML: 10 laps completed, pit-stop event with fuel added, no out-of-fuel stop, final position ≤ 4. The result folder is read-only to the agent and no shell is available, so results and car setups cannot be edited outside the game; input is keyboard only.
Proposed partial credit
30% pit stop made with fuel; 30% finished without running dry; 40% scaled by position
R13.01.01.03-T09 Complete a gravel rally stage with limited damage in BeamNG
Original instruction
BeamNG.drive rally scenario on a Utah gravel stage with pace notes in the HUD. Complete the stage under the target time 4:30 with total vehicle damage below $3,000 (scenario damage meter).
- Platforms
win· Windows- Applications
- BeamNG
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
rt· Real-time controlkey· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Scenario end report (harness extension): finish time, damage value. Pass at time ≤ 4:30 and damage < 3,000. The Lua console and world editor are disabled for the agent and BeamNGpy runs harness-side only, so damage and timing cannot be altered by script.
Proposed partial credit
30% stage finished; 40% scaled by time; 30% damage under limit
R13.01.01.03-T10 Win a HexGL anti-gravity race in the browser
Original instruction
HexGL (self-hosted WebGL racer) is open on the default track, difficulty Casual. Finish the 3-lap race with your shield not depleted and a total time under 2:30.
- Platforms
web· Any OS (web)- Applications
- racing games
- Interfaces
brow· Browser- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
rt· Real-time controlkey· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
JS bridge on the game object (closure-scoped, harness only): race finished, shield > 0, total time < 150 s. DevTools are disabled by browser policy, so game state cannot be patched by script; only trusted (isTrusted) keys are accepted.
Proposed partial credit
40% race finished; 20% shield intact; 40% scaled by time
R13.01.01.03-T11 Lap Monza under target time in Assetto Corsa without penalties
Original instruction
Assetto Corsa hotlap session at Monza in the Ferrari 458 GT2 with default setup. Set a valid lap under 1:51.0 within 15 minutes of wall-clock time, keyboard or the virtual gamepad.
- Platforms
win· Windows- Applications
- Assetto Corsa
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
rt· Real-time controlkey· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 5 / 5
Proposed verifier
Shared-memory telemetry captured by the harness: best valid lap time (track-limit invalidations excluded). Content Manager and in-game apps that change physics are disabled.
Proposed partial credit
20% valid lap; 80% scaled by best valid lap from 2:05 to 1:51
R13.01.01.03-T12 Reverse a semi-trailer into a loading dock in Euro Truck Simulator 2
Original instruction
Euro Truck Simulator 2 save starts at a depot with a job whose trailer must be delivered. Choose the 'hard' parking option and reverse the trailer into the marked dock bay with no collisions.
- Platforms
win· Windowsmac· macOSlin· Linux- Applications
- Euro Truck Simulator 2
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
rt· Real-time controlgeom· Precise geometryvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Telemetry SDK plugin (harness side) logs trailer pose and job events: delivery event with 'hard parking' flag, collision count 0, trailer damage unchanged. The developer console is disabled (g_developer 0, config read-only) and the agent has no shell access to the profile folder.
Proposed partial credit
30% trailer aligned to bay; 40% delivery accepted with hard parking; 30% no collisions
R13.01.01.03-T13 Hold a drift for a target score in BeamNG
Original instruction
BeamNG.drive drift scenario on the Gridmap with a rear-wheel-drive coupe. Accumulate a single uninterrupted drift chain scoring at least 5,000 points around the marked cones without spinning out.
- Platforms
win· Windows- Applications
- BeamNG
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
rt· Real-time controlkey· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Drift extension score log: max single chain ≥ 5,000 with no spin event (yaw > 180°) in that chain. The Lua console and world editor are disabled for the agent and BeamNGpy runs harness-side only, so score and vehicle state cannot be set by script.
Proposed partial credit
50% scaled by best chain score up to 2,500; 50% scaled from 2,500 to 5,000
R13.01.01.03-T14 Complete the SuperTuxKart tutorial
Original instruction
Start SuperTuxKart and play the built-in tutorial track to the end, following its prompts for acceleration, steering, items, nitro and skidding.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- SuperTuxKart
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
rt· Real-time controlkey· Shortcuts- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 1 / 5
Proposed verifier
Game config/achievements file: tutorial completion flag set during the run. Agent input via OS key events only; the game console and artist debug mode are disabled and the config folder is read-only to the agent, so the flag cannot be written directly.
Proposed partial credit
50% scaled by the first half of tutorial checkpoints passed; 50% scaled by the rest, full credit on the completion flag
R13.01.01.03-T15 Score five goals in Rocket League free play
Original instruction
Rocket League (offline Free Play, Octane, keyboard controls) on DFH Stadium. Hit the ball into the orange goal 5 times within 5 minutes.
- Platforms
win· Windows- Applications
- Rocket League
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
rt· Real-time controlkey· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
A harness BakkesMod plugin (read-only) logs goal events in free play; fallback: goal-explosion detection by frame classifier. Plugin console is unavailable to the agent.
Proposed partial credit
20% first goal; 20% second goal; 20% third goal; 40% fourth and fifth goals
R13.01.01.03-T16 Beat a ghost replay in a SuperTuxKart time trial
Original instruction
SuperTuxKart time trial on 'Zen Garden', 3 laps, with a provided ghost replay of 2:10.5 loaded. Finish faster than the ghost.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- SuperTuxKart
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
rt· Real-time controlkey· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 4 / 5
Proposed verifier
Result file: race time < 130.5 s and 3 laps. Ghost is visible for reference only. The game console and artist debug mode are disabled, input is OS key events only, and the result folder is read-only to the agent, so results cannot be edited.
Proposed partial credit
30% finished; 70% scaled by time from 2:40 to 2:10.5
R13.01.01.03-T17 Complete a Need for Speed lap with at most one collision
Original instruction
In DOSBox, The Need for Speed practice on 'City' track in the Lamborghini Diablo with traffic on. Complete one lap with at most one collision with traffic or walls.
- Platforms
mac· macOSwin· Windowslin· Linux- Applications
- racing games
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
rt· Real-time controlkey· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Collision counter from DOS memory (harness side), fallback frame-diff impact heuristic; lap completion from the lap counter. The DOSBox debugger and config file are not accessible to the agent, so DOS memory cannot be patched; input is OS key events only.
Proposed partial credit
40% lap completed; 60% collisions ≤ 1 (half credit if 2-3)
R13.01.01.03-T18 Reach 1,000 m in Hill Climb Racing on Android
Original instruction
On the Android emulator, Hill Climb Racing is on stage 'Countryside' with the starter jeep (no upgrades). Using the on-screen gas and brake pedals, drive as far as possible and reach 1,000 m before fuel runs out or the driver's head hits the ground.
- Platforms
and· Android- Applications
- Hill Climb Racing
- Interfaces
mob· Mobile touch- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G3· GUI-only- Interaction primitives
touch· Touch gesturesrt· Real-time controlvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Best distance read from the app's saved stats (pulled with adb by the harness) and cross-checked with a screen recording OCR of the end screen. adb and root are not exposed to the agent, so the saved stats cannot be edited; input comes only from emulator touches.
Proposed partial credit
50% scaled by best distance up to 500 m; 50% scaled from 500 to 1,000 m
R13.01.01.03-T19 Plough a field with full coverage in Farming Simulator
Original instruction
Farming Simulator 22 save: a tractor with a plough at the edge of field 7 (2.1 ha). Plough the whole field with manual steering (no AI helper, no GPS mod) so at least 95% of the field area is ploughed, without driving outside the field boundary.
- Platforms
win· Windowsmac· macOS- Applications
- Farming Simulator
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
rt· Real-time controlpath· Continuous pathnav· Pan/zoom/scroll- Work product
WP11· Game/sim state- Conditions
long· Long-horizon- Proposed difficulty
- 4 / 5
Proposed verifier
Savegame density maps parsed: ploughed fraction of field 7 ≥ 0.95; helper-hire events absent from the log. The developer console is disabled and the savegame folder is read-only to the agent (no shell), so density maps cannot be edited outside the game.
Proposed partial credit
50% ploughed fraction from 0.5 to 0.725; 50% from 0.725 to 0.95; score minus one fifth if a helper is used
R13.01.01.03-T20 Pass a moose-test lane change at 70 km/h in BeamNG
Original instruction
BeamNG.drive scenario: a cone-marked ISO 3888-2 avoidance course. Enter at 70 ±2 km/h, release the throttle at the entry gate and complete the double lane change without hitting cones.
- Platforms
win· Windows- Applications
- BeamNG
- Interfaces
desk· Desktop GUI- GUI need · task setting
G3· GUI-only- GUI need · ordinary access
G2· GUI-essential- Interaction primitives
rt· Real-time controlkey· Shortcutsvis· Visual judgment- Work product
WP11· Game/sim state- Conditions
- None declared
- Proposed difficulty
- 3 / 5
Proposed verifier
Scenario extension logs entry speed at the gate, throttle state and cone displacement: pass when speed in band, throttle released, and 0 cones moved. The Lua console and world editor are disabled for the agent and BeamNGpy runs harness-side only, so speed and cones cannot be set by script.
Proposed partial credit
30% entry speed in band; 20% throttle released; 50% no cones hit
Mapped research references
Existing benchmark-to-class links from the research catalog. They are not record-level annotations, executions or product endorsements.