Source-linked AI summary

One Thousand and One Hours: Self-driving Motion Prediction Dataset

John Houston, Guido Zuidhof, Luca Bergamini, Yawei Ye, Long Chen, Ashesh Jain, Sammy Omari, Vladimir Iglovikov, Peter Ondruska

arXiv:2006.14480v2cs.CVcs.LGcs.RO

TL;DR

Self-driving research lacks broadly available large-scale datasets for motion prediction, planning, and related downstream tasks. This paper releases a route-focused dataset with detailed perception outputs, semantic mapping, aerial context, software, and baselines. The authors report meaningful performance increases for motion forecasting and motion planning as dataset size grows.

  • Problem

    Large-scale datasets for motion prediction and planning are scarce, while planning is difficult to model and evaluate from real data.

  • Method

    The paper releases a dataset of traffic scenes with full perception outputs, a detailed semantic map, aerial imagery, software tools, and motion forecasting and planning baselines.

  • Results

    The dataset is the largest and most detailed available for prediction and planning, and performance increases meaningfully for both tasks with more training data.

  • Takeaways & Limitations

    The release broadens access to self-driving machine-learning development and suggests that still larger datasets may be desirable.

Abstract

from arXiv · show

Motivated by the impact of large-scale datasets on ML systems we present the largest self-driving dataset for motion prediction to date, containing over 1,000 hours of data. This was collected by a fleet of 20 autonomous vehicles along a fixed route in Palo Alto, California, over a four-month period. It consists of 170,000 scenes, where each scene is 25 seconds long and captures the perception output of the self-driving system, which encodes the precise positions and motions of nearby vehicles, cyclists, and pedestrians over time. On top of this, the dataset contains a high-definition semantic map with 15,242 labelled elements and a high-definition aerial view over the area. We show that using a dataset of this size dramatically improves performance for key self-driving problems. Combined with the provided software kit, this collection forms the largest and most detailed dataset to date for the development of self-driving machine learning tasks, such as motion forecasting, motion planning and simulation. The full dataset is available at http://level5.lyft.com/.

1 Introduction

Large-scale datasets have advanced self-driving perception, but downstream motion prediction, planning, and simulation require behavioural observations, interactions, and semantic maps that are less broadly available. This work addresses that gap with a large, detailed dataset focused on motion forecasting and planning.

  • Motivation: Downstream self-driving tasks require behavioural observations, interactions, and semantic maps in addition to perception outputs.These requirements support data-driven motion prediction, trajectory planning, and simulation.
  • Motivation: Broadly available datasets for motion prediction, planning, and simulation are much more limited than perception datasets.Existing downstream datasets are mostly available through large-scale industrial efforts using in-house data.
  • Contribution: The dataset targets motion forecasting and planning rather than training perception systems, using output from an already trained perception system.It processes large quantities of new data for motion prediction.
  • Dataset: 1,000 hours of traffic scenes capture traffic-participant motions around 20 self-driving vehicles driving over 26,000 km along a suburban route.The collection is presented as the largest dataset to date for motion prediction.
  • Dataset: The release includes over 15,000 human-annotated semantic-map elements, a 74 km2 aerial image at 6 cm per pixel, L5Kit, and baseline solutions.These components support access, visualization, motion forecasting, and motion planning development.

2 Related Work

Related work spans perception, prediction, and planning, but open datasets and evaluations remain especially limited for downstream tasks. The paper distinguishes its dataset through route-focused scale, richer perception outputs, detailed mapping, aerial context, and an open planning evaluation.

  • Self-driving stack: Self-driving stacks estimate nearby actors, predict their future motion and intent, and use those predictions to plan the vehicle trajectory.Figure 2 presents this sequence from raw sensor input through perception and prediction to planning.
  • Perception datasets: Perception datasets commonly provide 3D object labels, while the released dataset instead processes already trained perception output for motion prediction.Its main target is not training perception systems.
  • Prediction datasets: Prediction methods require detailed environmental information such as semantic maps, yet representative large-scale training datasets are rare and mostly proprietary.The most relevant existing open dataset, Argoverse Forecasting, provides 300 hours and a lightweight map encoding lane centers.
  • Prediction datasets: The dataset differs from Argoverse by providing 1,000 hours along a single route and full perception outputs including bounding boxes and class probabilities.The route focus is motivated by likely early deployment along a few high-demand ride-hailing routes.
  • Prediction datasets: Its semantic map contains more than 15,000 human annotations rather than only lane centers, and the release also provides a high-resolution aerial image.The aerial representation may support development of semantic-map-free solutions.
  • Planning datasets: Planning has received relatively little community attention because it is difficult to model and evaluate from real data, especially under closed-loop evaluation.Closed-loop evaluation requires collecting new data that are absent from the dataset.
  • Planning datasets: An imitation-learning planning baseline with perturbations enables the first open evaluation of this approach and improves with more training data.The approach had previously been demonstrated only on a proprietary dataset with no available open equivalent.

3 Dataset

The released dataset contains 170,000 annotated driving scenes, supported by HD semantic and aerial maps, and collected from a sensor-equipped self-driving fleet on a fixed route.

  • 170,000 scenes, each 25 seconds long, capture the self-driving vehicle, nearby traffic participants, and traffic-light states.
  • The collection uses seven cameras, three LiDARs, and five radars processed by a perception system to represent traffic participants’ motion.
  • The data was collected by a fleet driving along a fixed route and split into train, validation, and test sets using an 83–7–10% ratio.
  • The HD semantic map contains 15,242 labelled elements, including 8,505 lane segments, with precise lane geometry and connectivity.
  • The aerial map covers 74 km2 at 6 cm per pixel and provides spatial information that can aid motion prediction.

4 Development tools

The software toolkit supports efficient scene access, visualisation, rasterisation, and baseline implementations for motion forecasting and SDV motion planning.

  • L5Kit provides data loading, visualisation, and implementations for motion forecasting and SDV motion planning baselines.
  • Scenes can be sampled around either the SDV for planning or other traffic participants for predicting their motions.
  • BEV rasterisations can combine semantic-map and aerial-map information into images suitable for conventional machine-learning pipelines.
  • Motion forecasting baseline: The forecasting baseline predicts traffic participants’ future (x,y)-positions over a 5-second horizon using ResNet-50 and stacked BEV histories.
  • Motion forecasting baseline: Adding motion history and increasing training-set size improve open-loop displacement-error performance.
  • Motion planning baseline: The planning baseline executes SDV trajectories in closed-loop simulations where other participants follow logged behaviour and the SDV may diverge from the log.

5 Conclusion

The paper presents a large, detailed dataset for prediction and planning, reporting performance gains for both tasks as dataset size increases. It argues that continued scaling and public release can support further self-driving development.

  • The dataset is described as the largest and most detailed available for training prediction and planning solutions.
  • The authors report meaningful performance increases for both motion forecasting and motion planning relative to existing alternatives.
  • Performance continues increasing with training-data size, motivating future datasets with tens of thousands or millions of hours.
  • The authors frame publication as a step toward democratising self-driving application development and faster progress toward autonomy.
Loading 2006.14480v2…