Uploaded May 2025 | Updated September 2026, 2 hours ago
Future Leaders in Robotics and AI Seminar Series: Towards Safe and Aligned Embodied AI in the Era of Robotics Foundation Models
Online Seminar
Ran (Thomas) Tian
PhD Student
University of California Berkeley
Robotics foundation models have revolutionized how robots perceive environments, learn from people, and interact with them. By seeing massive amounts of data, these models can learn implicit representations of the world and complex daily tasks. Despite the remarkable progress, robots relying on these models don’t inherently become better at doing what humans prefer. Often, they become less inclined to act in accordance with human preferences because the objectives suitable for training these large-scale paradigms are only proxies for what we—the users and stakeholders—care about. By optimizing an incomplete or misspecified objective, these robotics models lead to undesirable behaviors at best and safety hazards at worst. In this talk, I will introduce our efforts to bring the success of preference alignment, widely adopted in non-embodied foundation models (e.g., large language models), to embodied contexts such as robot manipulation and autonomous driving, enabling robots to align their behavior with human preferences in the open world.
Future Leaders in Robotics and AI Seminar Series: Towards Safe and Aligned Embodied AI in the Era of Robotics Foundation Models
Online Seminar
Ran (Thomas) Tian
PhD Student
University of California Berkeley
Robotics foundation models have revolutionized how robots perceive environments, learn from people, and interact with them. By seeing massive amounts of data, these models can learn implicit representations of the world and complex daily tasks. Despite the remarkable progress, robots relying on these models don’t inherently become better at doing what humans prefer. Often, they become less inclined to act in accordance with human preferences because the objectives suitable for training these large-scale paradigms are only proxies for what we—the users and stakeholders—care about. By optimizing an incomplete or misspecified objective, these robotics models lead to undesirable behaviors at best and safety hazards at worst. In this talk, I will introduce our efforts to bring the success of preference alignment, widely adopted in non-embodied foundation models (e.g., large language models), to embodied contexts such as robot manipulation and autonomous driving, enabling robots to align their behavior with human preferences in the open world.
