🚘 Language2Drive

Do Vision-Language-Action Models Really Follow the Instructions in Autonomous Driving?

Same scene. Same seed. Different instructions.

👥 Authors

Zihan You^1,2,,†, Zehua Fan^3,, Wenqi Lyu^4,, Aoqi Wang^5,,
Ni Wang⁶, Yifei Yang⁷, Linge Zhao⁸, Huan Li⁹, Sining Ang¹⁰, Yuguang Yang¹,
Jueyu Chen^10,11, Xiangyu Li¹², Qianshan Wei¹³, Jieyuan Pei^1,14,15, Yuxin Wang¹⁶, Zhiwen Chen¹⁵,
Jiaqi Li¹⁷, Wenxuan Song¹⁵, Yanyuan Qiao¹⁸, Benjin Zhu¹⁹, Haoang Li¹⁵, Yan Wang^1,‡

🏛️ Affiliations

¹ THU · ² SEU · ³ SJTU · ⁴ Adelaide · ⁵ BISTU · ⁶ Amazon · ⁷ BJTU
⁸ WHU · ⁹ HUST · ¹⁰ USTC · ¹¹ NIO · ¹² UT Austin · ¹³ CAS
¹⁴ HiThink · ¹⁵ HKUST(GZ) · ¹⁶ AMD · ¹⁷ HKU · ¹⁸ EPFL · ¹⁹ Li Auto

_{* Core contributors   † Project Lead   ‡ Corresponding author}

4 instruction scopes · 3 runtime groups · 5 evaluation backends

🧭 Benchmark in 30 seconds

Language2Drive evaluates whether changing a natural-language instruction changes the feasible behavior selected and executed by a driving VLA. Its Paired Counterfactual Unit (PCU) fixes the initial driving scene and changes only the instruction, separating individual task completion from completion of both counterfactual alternatives.

• Matched pairs: same scene, initial state, non-language inputs and simulator seed; different mutually exclusive instructions.
• Four scopes: atomic actions, persistent behaviors, scene-grounded navigation and long-horizon instruction programs.
• Five backends: CARLA and RealSim closed loop; nuScenes, NAVSIM and WOD-E2E open loop.
• Training-free evaluation: all reported baseline systems use the training-free protocol.
• Adapt once: the unified SDK connects a model to all five native evaluators through one trajectory interface.

🔀 PCU: the core evaluation unit

Shared across the pair	Changed between the arms	Evaluated
Scene, initial state, non-language inputs and seed	Active natural-language instruction	Requested behavior and rejection of the opposite contract
Checkpoint and observation configuration	Feasible target behavior	Native safety, progress, deadline and stability requirements

At a shared junction, one arm can request “Turn left” and the other “Go straight.” Each arm starts independently with reset model memory. Strict paired success requires completing both requests. The model receives public observations and the active instruction; geometric targets, evaluator contracts and the opposite-arm instruction remain with the evaluator.

Scope	What the instruction requires	Example task
A · Atomic action	Select a local maneuver	Turn, change lane or park
B · Persistent behavior	Maintain a constraint over time	Speed band, lane choice or following behavior
C · Scene-grounded navigation	Resolve a destination or relation in the observed scene	Parking bay, named landmark or obstacle-relative destination
D · Long-horizon instruction program	Execute stages in the specified order	Conditional destinations and multi-stage maneuvers

🏗️ Benchmark construction

Source scenes are grounded geometrically, paired with feasible instructions, compiled into executable contracts and validated before evaluation. Paraphrases test sensitivity to equivalent wording.

🛣️ Benchmark Tracks

Track	Evaluation	Manuscript / frozen scope	Model output	Native guide
CARLA-L2D	Closed loop; six purpose-designed maps	64 paper PCUs, A–D; 384 core rollouts/system, three expressions	Ego XY → declared controller	CARLA
RealSim	Closed loop; AlpaSim / NuRec	30 PCUs / 23 scenes, A–B; 60 slots/system, seed 1701	6 s / 60 points, 10 Hz	RealSim
nuScenes	Open loop	75 release PCUs: 15 dev + 60 held-out; paper uses held-out canonical	3 s / 12 points, 4 Hz	nuScenes
NAVSIM	Open loop	38 PCUs / 19 recordings, four wordings	4 s / 8 points, 2 Hz	NAVSIM
WOD-E2E	Open loop	241 P0 scored PCUs: 76 STOP/GO + 165 speed-choice	5 s / 20 points, 4 Hz	WOD-E2E

Scopes are instruction categories; tracks are evaluation environments. Counts are not interchangeable. CARLA's source inventory is broader than the 64-PCU paper cohort; record tier, split and seed. The SDK defaults to base and evaluates dev/test separately. CARLA PSM uses 24 additional routes and requires the native diagnostic evaluation. WOD's public workload can include construction candidates beyond the frozen 241-PCU scored subset.

📦 Resources in this repository

This Hugging Face repository hosts benchmark workload definitions, replay bundles and runtime resources. The official versioned dataset and model/checkpoint releases are coming soon.

Resource group	Browse / download	Contents
CARLA-L2D	CARLA resources	Benchmark and runtime archives, manifests, checksums and integration sources
RealSim	RealSim resources	PCUs, public workload, run plan, evaluator sidecar and runtime configuration
nuScenes	nuScenes resources	Asset manifest and nuscenes-paper-1.0-replay.zip
NAVSIM / WOD-E2E	NAVSIM code · WOD-E2E code	Native evaluator guides, source-data requirements and release profiles

The historical RealSim directory is spelled langauge2drive.RealSim; use that exact prefix when downloading.

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Mike087/Language2Drive",
    repo_type="dataset",
    local_dir="./language2drive-resources",
    allow_patterns=["langauge2drive.RealSim/**"],
)

Select language2drive.carla/** or language2drive.OpenloopTrack.nScenes/** for the other available resource groups. Repository authorization uses your Hugging Face account. Model checkpoints are obtained from their original publishers. Source scenes follow the access terms of their upstream datasets.

🗂️ Dataset / Evaluation Format

Layer	Contents	Consumer
Public observation	Cameras, observed ego state/history and active instruction	Model adapter
Trajectory request	Track, ego frame, reference point, timestamp grid and model seed	Model adapter
Evaluator metadata	PCU/arm identity, contracts, pair binding and expected inventory	Track runner / scorer
Submission	Native trajectory records	Native scorer
Result	Scores, scheduled coverage, provenance and protocol	Leaderboard / analysis

The SDK accepts +x forward, +y left, metres, with explicit frame, reference point and timestamps. For example, a NAVSIM output uses the rear axle and eight points at 2 Hz:

{
  "coordinate_frame": "ego_at_anchor",
  "reference_point": "rear_axle",
  "timestamps_s": [0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0],
  "xy_m": [[1, 0], [2, 0], [3, 0], [4, 0], [5, 0], [6, 0], [7, 0], [8, 0]]
}

nuScenes uses ego_pose_origin; RealSim uses rig_origin. The adapter guide documents each output contract, and the protocol specifies native submission files and cohort identities.

🏆 Main Results / Leaderboard

All reported systems use the training-free protocol. Values below are transcribed from the supplied manuscript; all scores are 0–100, higher is better. Systems are compared within each track and profile. Project-page leaderboard source · Machine-readable results · Submission guide.

Click a model name to read its paper. Explore the interactive metric guide for paired success, graded compliance and the L2D score.

📊 CARLA-L2D

paper: 64 PCUs · 3 expressions · seed 1.

System	SISR	CFPS	SemISR	Sem-CFPS	SCS	PCS	PSM	DrivingScore	L2D
Alpamayo 1.5 · model card	8.07	1.04	42.19	30.73	26.18	12.50	70.10	62.69	12.20
SimLingo	7.81	1.04	48.70	35.94	23.51	11.94	62.30	65.04	12.15
LMDrive	12.24	0.00	29.17	18.23	21.45	9.26	88.84	50.12	10.39
OneVL	0.52	0.00	25.00	3.65	16.69	7.05	77.72	50.69	4.87
Qwen-Drive-1.0-4B	1.82	0.00	15.62	3.12	17.05	7.09	76.68	43.67	4.31
Impromptu-VLA	0.26	0.00	19.79	6.25	14.48	4.61	77.43	42.02	3.87

📊 RealSim

realsim-paper-1.0: 30 PCUs · seed 1701.

System	SISR	CFPS	SemISR	Sem-CFPS	SCS	PCS	Safety	DrivingScore	L2D
Alpamayo2-Super · model card	13.33	0.00	18.33	0.00	21.43	0.00	93.33	97.10	9.22
Alpamayo 1.5 · model card	11.67	0.00	15.00	0.00	18.12	1.11	100.00	97.17	8.38
SimLingo	6.67	3.33	13.33	6.67	16.67	10.00	68.33	83.09	7.58
AutoVLA	1.67	0.00	8.33	6.67	9.73	6.67	41.67	76.94	3.36
Qwen-Drive-1.0-4B	0.00	0.00	1.67	0.00	4.44	0.00	43.33	69.16	0.92
Impromptu-VLA	0.00	0.00	1.67	0.00	3.71	0.00	30.00	68.85	0.78
OneVL	0.00	0.00	0.00	0.00	0.00	0.00	43.33	43.82	0.44
ReCogDrive	0.00	0.00	0.00	0.00	0.00	0.00	33.33	43.57	0.38

📊 nuScenes

v1.0-test · 60 held-out PCUs · canonical.

System	SemSISR	SemCFPS	SCS	PCS
Impromptu-VLA	15.83	0.00	20.21	0.00
UniDriveVLA	16.67	0.00	25.00	0.00
OpenDriveVLA-0.5B	13.33	3.33	19.06	4.30
OmniDrive	7.50	0.00	29.24	8.61

📊 NAVSIM

paper historical scorer · 38 PCUs · 4 wordings.

System	SemSISR	SemCFPS	SCS	PCS
Qwen-Drive-1.0-4B	0.000	0.000	5.472	0.000
AutoVLA	3.947	0.000	8.172	0.000
DriveMA-2B	8.553	0.000	11.244	2.242
ReCogDrive	0.000	0.000	4.400	0.000
OneVL	3.618	0.000	8.428	0.000

📊 WOD-E2E

P0 · 241 PCUs · canonical.

System	SemSISR	SemCFPS	SCS	PCS
DriveMA-2B	38.38	1.24	49.33	7.88
Fast-dDrive	34.85	0.00	46.42	5.27
Qwen-Drive-SFT	34.44	0.00	50.58	9.37
Qwen-Drive-RL	35.06	0.00	49.58	7.58

The CARLA table includes the six principal systems; the paper's Senna high-level bridge is a separate diagnostic configuration. NAVSIM numbers use the historical manuscript scorer. New SDK runs use score_corrected.py; keep corrected evaluations in a separate leaderboard profile. Full protocol, coverage and failure denominators must accompany submissions. Rankings follow each track’s native metrics and evaluation profile.

📐 Reading the metrics

SISR / CFPS measure strict individual / paired completion. SCS / PCS measure graded individual / paired compliance. Closed-loop SemISR / Sem-CFPS are relaxed proxies; open-loop SemSISR / SemCFPS apply exclusive semantic success. Wording robustness is evaluated using equivalent paraphrases. L2D combines the native language and driving score families with the track's nonlinear aggregation. Full definitions are in the evaluation protocol.

🚀 Evaluate your model

Install the coordinator from the source repository or a release wheel, select your runtime groups, and implement one model adapter:

git clone https://github.com/ZihanYou/Language2Drive.git
cd Language2Drive
python -m pip install -e '.[openloop]'
language2drive init --workspace ./l2d-workspace
language2drive setup --workspace ./l2d-workspace --tracks all --runtime --download

setup selects official Hugging Face / PyPI endpoints when reachable and mirror / Tsinghua endpoints otherwise. Authenticated benchmark downloads use the official Hugging Face endpoint. CARLA and RealSim use separate simulator environments. Licensed source data, GPU services and model checkpoints are configured according to the selected backend.

• Quick start and all five native reproduction commands
• Full GPU evaluation guide
• Installation and resource setup
• Leaderboard submission guide
• Interactive project preview, paired cases and demonstration film

🗓️ Release Roadmap

Status	Deliverable
✅	Benchmark source, five native evaluators and a unified model-adapter SDK
✅	Wheel build and selectable environment/resource setup
✅	Training-free baseline tables and paired case visualizations
✅	Evaluation resources in this Hugging Face repository
⏳	Official versioned dataset release
⏳	Official model and checkpoint release

📚 Citation

If Language2Drive supports your research, please cite our paper.

@misc{language2drive,
  title = {Language2Drive: Do Vision-Language-Action Models Really Follow the Instructions in Autonomous Driving?},
  author = {You, Zihan and
           Fan, Zehua and
           Lyu, Wenqi and
           Wang, Aoqi and
           Wang, Ni and
           Yang, Yifei and
           Zhao, Linge and
           Li, Huan and
           Ang, Sining and
           Yang, Yuguang and
           Chen, Jueyu and
           Li, Xiangyu and
           Wei, Qianshan and
           Pei, Jieyuan and
           Wang, Yuxin and
           Chen, Zhiwen and
           Li, Jiaqi and
           Song, Wenxuan and
           Qiao, Yanyuan and
           Zhu, Benjin and
           Li, Haoang and
           Wang, Yan},
  howpublished = {Research manuscript},
  url = {https://github.com/ZihanYou/Language2Drive}
}

📄 License

Original Language2Drive code uses the MIT License. Dataset, simulator, runtime and third-party integration resources retain their respective licenses. See the CARLA resource license, third-party notices and RealSim notices. This card preserves the repository's component-specific license designation. Manuscript figures are research materials and are separate from the code license.