Last Update: 03.12.2024

Status: OPEN

Type: bachelors/masters (short/long)

Contact: Florian Bolli bollif@ethz.ch and Prof. Tobi Delbruck (tobi@ini.uzh.ch)

___________________________________________

We are using our INIvincible RoboRacer car to develop and test ideas about adaptive nonlinear neural control. PhD student Florian Bolli leads the team and carries out experiments with Marcin Paluch and several bachelor and masters students. The INIvincible teams cooperates with the ETH PBL ForzaETH team and does experiments every Monday in the Dubendorf hangar track area.

Open Projects

You shall not pass: Vision-Based Defensive Racing for Autonomous Vehicles

In this project, you will develop and implement a defensive racing system for our RoboRacer 1/10 scale autonomous car. By equipping the car with a rear-view camera and/or additional sensors, you will enable it to detect approaching opponents, estimate their position and velocity, and strategically defend the racing position against a faster opponent.

The ultimate goal is to combine perception, opponent state estimation, and autonomous control to prevent faster opponents from overtaking — first in simulation and ultimately on our physical race car.

Your Tasks

• Defensive racing strategies: Develop defensive behaviors using our existing Model Predictive Control (MPC) or Reinforcement Learning  frameworks.

• Rear-view perception: Equip our RoboRacer with a rear-facing camera (or other sensors) and train a neural network to detect approaching opponents.

• Distance measurement: Investigate and integrate suitable sensors to measure the distance to opponents behind the vehicle.

• Opponent state estimation: Combine sensor measurements to estimate the opponent's relative position, velocity, and potentially its intended overtaking maneuver.

• Simulation and real-world testing: Evaluate your algorithms in simulation before deploying and testing them on our physical car.

What Makes This Project Exciting?

• Hands-on robotics: Work directly with our small-scale autonomous racing vehicles, combining hardware, sensors, and software.

• Machine learning meets control: Apply neural networks, state estimation, and advanced control algorithms to a challenging real-world problem.

• Pragmatic engineering: Solve practical problems in a complex robotic system, from sensor integration and debugging to real-time performance optimization.

• From simulation to the racetrack: See your algorithms come to life on a physical vehicle and compete against other autonomous race cars.

What You'll Learn

This project offers practical experience in computer vision, sensor fusion, autonomous vehicle control, and robotics software development. You will work within an existing autonomous racing framework and have the freedom to investigate and implement your own solutions.

Ideal Background

• Interest in robotics, autonomous systems, and machine learning.

• Basic programming experience in Python.

• Familiarity with ROS 2, computer vision, or control theory is beneficial but not required.

• Motivation to work with physical hardware and solve challenging engineering problems.

The challenge: Can you make a slower autonomous car clever enough to defend its position against a faster opponent?

Literature:

RPGD: A Small-Batch Parallel Gradient Descent Optimizer with Explorative Resampling for Nonlinear Model Predictive Control

https://ieeexplore.ieee.org/document/10161233

Deep Reinforcement-Learning-Guided Model Predictive Control for Preventing Overtakes in Autonomous Racing

https://arxiv.org/abs/2607.20973

Last Update: 03.12.2024

Status: OPEN

Type: bachelors/masters (short/long)

Contact: Florian Bolli bollif@ini.uzh.ch and Prof. Tobi Delbruck (tobi@ini.uzh.ch)

___________________________________________

We are using our INIvincible RoboRacer car to develop and test ideas about adaptive nonlinear neural control. PhD student Florian Bolli leads the team and carries out experiments with Marcin Paluch and several bachelor and masters students. The INIvincible teams cooperates with the ETH PBL ForzaETH team and does experiments every Monday in the Dubendorf hangar track area.

Work In Progress Projects

Opponent detection and overtaking

Background: F1TENTH is about racing

The races have 2 categories:

• Time Trials:

  • car is alone on the track

  • count number of consecutive laps within 5 mins

  • fastest lap

• Head2Head racing

  • 1 opponent ( maybe more )

  • Who completes 10 laps first 

Problem: Our sytstem has no idea about overtaking. Not even about opponent detection.

What we already have:

• lidar measurements

• Information about the track border

• Algorithms that can avoid obstacles / pass opponents

• MPC avoiding scanned obstacles

• Alternative raceline strategy 

Goals:

• Implement a detection system that can estimate the positions and states of opponents

• Explore some common overtaking strategies and implement some of them

• Train RL agents to race each other

• Implement virtual opponents on the pysical car ( for RL overtaking )

Papers:

α-RACER: Real-Time Algorithm for Game-Theoretic Motion Planning and Control in Autonomous Racing using Near-Potential Function https://arxiv.org/abs/2412.08855

Predictive Spliner: Data-Driven Overtaking in Autonomous Racing Using Opponent Trajectory Prediction: https://arxiv.org/abs/2410.04868

M-Predictive Spliner: Enabling Spatiotemporal Multi-Opponent Overtaking for Autonomous Racing: https://arxiv.org/abs/2506.16301