Access nuReasoning
Please provide the requested information and review the nuReasoning Dataset License and Terms of Use before requesting access.
Please provide your full name, email address, and full institution or mcompany name. Avoid abbreviations where possible and ensure that the information you provide is accurate.
nuReasoning is a reasoning-centric multimodal autonomous driving dataset provided for research and development subject to the applicable nuReasoning Dataset License and Terms of Use.
Log in or Sign Up to review the conditions and access this dataset content.
nuReasoning
nuReasoning is a reasoning-centric multimodal autonomous driving dataset for evaluating and training end-to-end driving systems in long-tail real-world scenarios.
Each sample is built around a driving clip with synchronized multi-camera images, LiDAR data, ego state, object annotations, HD map, routing, and frame-level reasoning annotations.
Reasoning annotations are organized into three complementary fields:
- Spatial Reasoning: structured scene context, 2D and 3D detections, object relations, map projections, ego-frame object future states, and potential future conflicts.
- Decision Reasoning: scene interpretation, critical components, driving decisions, and causal reasoning traces.
- Counterfactual Reasoning: alternative actions, risk levels, safety outcomes, and explanations of why unsafe or suboptimal actions should be avoided.
Release Plan: 60% of the nuReasoning dataset (including 10K training clips, 1K validation clips, and 1K test clips) has been released for the nuReasoning Challenge 2026. The remaining data will be released after the challenge concludes at the end of 2026.
Dataset Summary
- Dataset name: nuReasoning
- Data source: Motional internal AV fleet
- Dataset owner: Motional AD Inc.
- Scenario focus: long-tail autonomous driving scenarios
- Dataset scale: 20K clips, approximately 105 hours of curated driving data
- Data splits: 17K train clips, 2K validation clips, and 1K private test clips
- Clip format: 20-second driving clips
- Sampling rate: 10 Hz sensor and state data
- Modalities: multi-view cameras, LiDAR, HD maps, ego state, object annotations, traffic-light states, route commands, and reasoning annotations
- Reasoning types: Spatial, Decision, and Counterfactual
Intended Use
nuReasoning is intended for research and development in reasoning-centric autonomous driving, especially for long-tail real-world scenarios that require spatial understanding, decision reasoning, counterfactual safety assessment, and planning evaluation.
Potential applications include:
- Evaluating whether VLMs, VLA models, and end-to-end driving models can reason over long-tail driving scenes and connect perception, map context, agent motion, and route intent to driving decisions.
- Supervised fine-tuning of models that consume multi-view driving context and produce structured spatial, decision, and counterfactual reasoning annotations.
- Benchmarking driving-specific reasoning across geometry, motion, decision-making, and counterfactual risk assessment tasks.
- Studying how reasoning supervision affects downstream planning performance in autonomous driving models.
This dataset should not be used to identify people, infer sensitive attributes, perform surveillance, or develop systems that violate applicable laws, privacy obligations, licensing terms, or safety requirements. It should also not be used as the sole basis for real-world autonomous driving deployment.
Data Collection and Labeling
nuReasoning is constructed from real-world driving logs. Candidate clips are selected to emphasize challenging and long-tail driving scenarios, such as vulnerable road users, work zones, stationary obstructions, traffic-control anomalies, emergency vehicles, animals on the road, and other safety-relevant situations.
The annotation process combines automatic scene extraction with verified reasoning annotations. Reasoning annotations are generated from synchronized sensor and state context, then reviewed and corrected as needed.
Dataset Structure
The dataset is organized as self-contained clip directories:
data/
├── train/
│ ├── part_1/
│ │ └── <log_name>_<keyframe_token>/
│ │ ├── metadata.json
│ │ ├── map.pkl
│ │ ├── ego_state/
│ │ │ └── <timestamp_us>.pkl
│ │ ├── annotations/
│ │ │ └── <timestamp_us>.pkl
│ │ ├── reasoning/
│ │ │ └── <timestamp_us>.json
│ │ └── ... sensor assets referenced by metadata.json
│ ├── part_2/
│ │ └── <log_name>_<keyframe_token>/
│ │ └── ...
│ └── ...
├── validation/
│ └── part_1/
│ └── <log_name>_<keyframe_token>/
│ └── ...
└── test/
└── <log_name>_<keyframe_token>/
└── ...
Each clip directory contains all scenario-level metadata needed to locate frame-level ego states, sensor assets, annotations, map context, mission goals, and reasoning files.
Clip Metadata Schema
Each metadata.json file contains clip-level information:
{
"clip_token": "string",
"clip_location": "string",
"log_name": "string",
"scenario_type": "string",
"start_timestamp_us": 0,
"end_timestamp_us": 0,
"frame_rate_hz": 10,
"total_frames": 201,
"map_annotation": "map.pkl",
"camera_calibrations": {},
"ego_dimensions": {},
"frames": []
}
Each frame record contains timestamp-aligned sensor, state, annotation, reasoning, and route information:
{
"token": "string",
"frame_index": 0,
"timestamp_us": 0,
"relative_time_s": 0.0,
"ego_state": "ego_state/<timestamp_us>.pkl",
"sensors": {
"cameras": {
"front": "path/to/image",
"front_left": "path/to/image",
"front_right": "path/to/image",
"left": "path/to/image",
"right": "path/to/image",
"back": "path/to/image",
"back_left": "path/to/image",
"back_right": "path/to/image"
},
"lidar": {
"point_cloud_path": "path/to/lidar"
}
},
"annotations": "annotations/<timestamp_us>.pkl",
"reasoning": "reasoning/<timestamp_us>.json",
"mission_goal": {
"command": "GO_STRAIGHT",
"route_path": []
}
}
Reasoning Annotation Schema
Reasoning annotations are stored in reasoning/<timestamp_us>.json. Each file is aligned to a frame in metadata.json.
{
"frame_index": 0,
"Spatial": {
"Summary": {},
"Per_camera_results": {},
"Object_relations": [],
"Map": {}
},
"Driving": {
"Scene description": "",
"Critical components": {},
"Driving decision": {
"Longitudinal": "",
"Lateral": ""
},
"Reasoning trace": ""
},
"Counterfactual": {
"Alternative actions": [],
"Top safety-critical actions": []
}
}
Spatial Reasoning
The Spatial field may include:
- Per-camera object detections
- 2D bounding boxes
- Object track tokens
- Cross-view and 2D-to-3D object associations
- Ego-frame 3D object positions and velocities
- Object relations relative to the ego vehicle
- Projected map features such as baseline paths, lanes, and crosswalks
- Future trajectories and potential conflict indicators
Decision Reasoning
The Driving field provides high-level driving interpretation and action-grounded reasoning, including:
- Scene description
- Critical agents and environmental factors
- Longitudinal and lateral driving decisions
- Concise causal reasoning traces linking scene context, route intent, and selected ego behavior
Example decision fields include:
{
"Driving decision": {
"Longitudinal": "Slow down gently",
"Lateral": "No lateral action"
},
"Reasoning trace": "Slow down gently to maintain a safe following distance to the decelerating vehicle ahead while keeping centered in the lane."
}
Counterfactual Reasoning
The Counterfactual field evaluates plausible alternative actions under the same scene context. Alternatives may be categorized as:
SafeSuboptimalUnsafe
Counterfactual annotations identify safety-critical outcomes such as collisions, route deviations, traffic-rule violations, unnecessary blocking, or increased risk to vulnerable road users.
Supporting Files
map.pkl: static local map context around the scenario, including lanes, baseline paths, lane boundaries, crosswalks, intersections, stop polygons, road blocks, traffic lights, and lane connectors.ego_state/<timestamp_us>.pkl: ego pose, velocity, acceleration, dimensions, history trajectory, and future trajectory.annotations/<timestamp_us>.pkl: object annotations and traffic-light states for the frame.- Sensor assets: paths to eight camera views and LiDAR data referenced by
metadata.json.
Loading the Dataset
Please use the nuReasoning DevKit to load and process the dataset.
Ethical Considerations
Autonomous driving datasets may contain sensitive scene information even after privacy processing. Users should:
- Respect all dataset license terms and access restrictions.
- Avoid attempts to identify people, vehicles, locations, or private attributes.
- Evaluate downstream models for safety, bias, robustness, and misuse risks.
- Treat model outputs as research artifacts unless they have passed the safety validation required for deployment.
- Use the dataset only in applications consistent with applicable laws, privacy obligations, and autonomous-driving safety practices.
License and Terms of Use
nuReasoning has separate terms for non-commercial and commercial use.
- Non-commercial use: See LICENSE-NONCOMMERCIAL.md.
- Commercial use: See LICENSE-COMMERCIAL.md. Commercial use requires a commercial license; contact nuScenes@motional.com.
Users are responsible for complying with the applicable license terms, institutional policies, and legal requirements.
- Downloads last month
- 108