Tuesday, November 19, 2024, 10:15 – 11:45am
ML Engineers from the Tesla AI team will be sharing exclusive content about building physical AI
► Talk 1
World Model for End-to-End Simulation and Evaluation
Tesla Autopilot engineers will share some recent work in generative modeling for high-fidelity multi-view video generation in simulation and model evaluation. One of the end-to-end driving model's biggest challenges is evaluating the system's performance in axes of both comfort and safety, since many undesirable behaviors are only observed in closed-loop controls. To reduce the gap between open-loop metrics and close-loop performance, we introduce a controllable, close-loop simulation system based on large-scale pre-trained video generative model that enables scalable model evaluation prior to deployment. We will demonstrate the system's capabilities under various driving conditions and challenges we encountered in practice, and how your research can make an impact in advancing the frontier of end-to-end simulation.
— Zhichun (Eric) Huang earned his Master’s degree in Machine Learning from CMU in 2021, where he collaborated with Prof. Zico Kolter on applications of deep equilibrium models. He joined Tesla Autopilot in 2022, leading efforts in scaling generative models for perception, end-to-end control and simulation.
► Talk 2
Scaling Generative Models for Decision Making Dhaval Shroff — Senior Staff Autopilot Machine Learning Engineer
* Rohan Dhesikan* — Senior Autopilot Software Engineer – Autopilot
Tesla’s Autopilot engineers will share some recent work ingenerative modeling for high-fidelity multi-view video generation in simulation and model evaluation. One of the end-to-end driving model's biggest challenges is evaluating the system's performance in axes of both comfort and safety, since many undesirable behaviors are only observed in closed-loop controls. To reduce the gap between open-loop metrics and close-loop performance, we introduce a controllable, close-loop simulation system based on large-scale pre-trained video generative model that enables scalable model evaluation prior to deployment. We will demonstrate the system's capabilities under various driving conditions and challenges we encountered in practice, and how your research can make an impact in advancing the frontier of end-to-end simulation.
— Dhaval Shroff earned his Master's in Robotics Systems Development from CMU, focusing on classical computer vision and multi agent path planning. He has worked at Tesla on the Autopilot/ AI team for over 10 years and currently trains large models that predict end-to-end learned policies from video.
— Rohan Dhesikan earned both his undergraduate and Master's degrees in Electrical and Computer Engineering from CMU. He conducted research at the Robotics Institute, developing planning algorithms for aerial exploration of subterranean environments. After graduating, he joined Tesla's Autopilot team, where he focuses on training, optimization, and compression of large models for end-to-end self-driving systems.
► Talk 3
From Driver-Assist to Driver-Less Theo Guerin — Staff Autopilot Machine Learning Engineer - Autopilot
* Amy Lin* — Autopilot Machine Learning Engineer - Autopilot
In this talk, Autopilot engineers will share the evolution of autonomy at Tesla, with an emphasis on high-precision products such as AutoPark and Actually Smart Summon. We will discuss the motivations behind and the challenges of training deep neural networks for real-time 3D perception and vehicle control.
— Amy Lin earned her undergraduate degree in Computer Science at CMU, where she conducted research in using deep learning for 3D reconstruction and pose estimation. She joined Autopilot's Fleet Learning team shortly after graduating, where she works on the high fidelity occupancy networks.
— Theo Guerin received his Bachelor of Science from Brown University in 2021, where he specialized in robotics. After two summer internships in the self-driving space and upon graduation, Theo joined the Motion Planning and Controls team in Autopilot. His work has spanned the planning and controls stack and most recently has been focused on the training and application of end-to-end models to driverless features like Actually Smart Summon.
Event Type: Career Presentation
Room Number: In Person
Building: ASA Conference Room, Gates Hillman 6115
Speaker's Name: ENGINEERING BREAKOUT SESSIONS - Deep Dive Technical Talks
Event Poster Title: Posters - All Tesla Presentations
Event Poster URL: www.cs.cmu.edu…
Event Website Title: SCS Careers
Event Website URL: www.cs.cmu.edu…