Research Engineer Internship at Avride in Austin, TX

Summary: Avride is a US-based developer of autonomous vehicles and delivery robots. The Research Engineer Internship focuses on using large datasets to train models and develop algorithms for autonomous vehicles, working closely with the ML Prediction and Planning team to enhance driving performance through innovative research projects. Responsibilities: - Take ownership of a research project focused on exploring how model ensembling strategies influence the gap between open-loop (training) and closed-loop (simulation) performance - Review relevant literature, formulate hypotheses, and prototype solutions using Python and ML frameworks (like PyTorch) - Implement and evaluate multiple ensembling approaches, including blending models trained with different random seeds, combining checkpoints from different training stages, and applying weighted averaging or learned blending of model outputs - Systematically compare single-model vs ensemble performance and seed diversity vs checkpoint diversity, and measure their impact on open-loop metrics (training/validation loss, accuracy) and closed-loop metrics (simulation performance, safety, stability) - Investigate the correlation (or lack thereof) between open-loop and closed-loop improvements, identify cases where ensembling improves one metric but degrades the other, and formulate hypotheses explaining the observed behavior - Evaluate and improve the behavior of ML-driven traffic agents in our autonomous driving simulator - Design evaluation functions that select trajectories with desired properties — from realistic to adversarial — and build quantitative metrics to measure how agent behavior changes - Run experiments on large scenario pools, and produce results that directly influence the team's roadmap for agent simulation - Work alongside your mentor to design, test, and iterate algorithms that select agent trajectories optimizing for different objectives: aggressiveness, interaction density, route fidelity - Build evaluation metrics for comparing agent behavior strategies: interaction intensity (time-to-collision, proximity), kinematics plausibility (acceleration, jerk), and distributional similarity to real traffic - Run experiments on large-scale scenario pools, comparing ML agents against baseline approaches and measuring the impact of different strategies - Work with production codebase: the prediction models you'll experiment with are the same ones deployed in our autonomous vehicles - Conclude your internship by presenting your methodology, experimental results, and data-driven recommendations on where trajectory ranking is sufficient and where model-level changes are required Required Qualifications: - Currently pursuing a Master's or PhD (highly preferred) in Computer Science, Robotics, Machine Learning, Applied Mathematics, or a related field with an expected graduation date between Winter 2026 and Spring 2027 - Strong understanding of deep learning, reinforcement learning, computer vision, optimization, or probabilistic modeling - Proficiency in Python and deep learning frameworks (PyTorch, TensorFlow) - Basic familiarity or willingness to learn C++ - Ability to read, understand, and implement algorithms from academic research papers - A strong analytical mindset for designing experiments and interpreting data - Highly collaborative, open to feedback, and excited to tackle unsolved problems in the autonomous driving space - Candidates are required to be authorized to work in the U.S - The employer is not offering relocation sponsorship, and remote work options are not available Required Skills: Machine Learning, Deep Learning, Reinforcement Learning, Computer Vision, Optimization, Probabilistic Modeling, Python, PyTorch, TensorFlow, C++, Research Paper Implementation, Experiment Design, Data Analysis, Openness to feedback Benefits: 1:1 Mentorship: Direct guidance from leading researchers and engineers in the autonomous vehicle industry to help you navigate technical roadblocks and grow your career., Massive Compute & Data: Access to state-of-the-art driving data to fuel your experiments., Networking & Culture: Invitations to tech talks, paper reading groups, intern social events, and cross-team collaborations.