AdaComp, National University of Singapore
Are you interested in Autonomous Driving? Do you want to be the brain of the next generation of smart cars?
To drive safely and smoothly in such traffic, the smart cars have to reason about and predict the motion of other nearby vehicles and pedestrians. The long-term goal of Adacomp Laboratory is to enable fluid human-robot interaction, collaboration, and ultimately co-existence. One important application is autonomous driving, especially, in heavy traffic. See the video below for an example of our recent work.
summit.mp4
We have identified motion prediction as a key challenge of autonomous driving: predicting the future positions of surrounding traffic participants, so that the robot vehicle can decide on its own motion accordingly.
There are two general approaches to motion prediction:
- Analytical prediction methods follow a set of handcrafted rules or optimize a handcrafted objective.
- data-driven methods learn prediction models from trajectory datasets.
They differ significantly in prediction accuracy, computational efficiency, data efficiency for learning, …, and each has strengths and weaknesses. The lack of systematic evaluation and comparison hinders the progress of autonomous driving. The two common evaluation metrics, average displacement error (ADE) and final displacement error (FDE), seem natural, but may not directly impact driving decisions and performance.
In this project, we will build an experimental framework (evaluation criteria, datasets, and a simulator) for systematic evaluation of motion prediction methods. We intend to release the experimental framework as open-source software and publish a paper on the findings to report on the state of the art.
You will gain experiences in some of the following:
- Choose representative analytical and data-driven models and conduct a comprehensive study on their performance using a richer set of evaluation metrics.
- Evaluate the prediction models for driving performance in a high-fidelity driving simulator.
- Evaluate the prediction models using both real-world datasets, e.g., ArgoVerse, and a real-time driving simulator, e,g, SUMMIT, to compare performance in “canned” and interactive settings.
- Propose new ideas to improve the prediction and planning methods for autonomous driving.
The expected duration of the project is 3 months, which may be extended based on mutual agreement. You may participate either on-site or remotely.
You are invited to submit the application along with your resume by emailing to moped. We expect you to have
- some basic knowledge in machine learning and prediction models
- strong software engineering skills Look forward to working with you on the exciting research!