Dissertation Title: Efficient and Generalizable Robot Learning via Physical, Geometric, and Semantic Priors

Date: 2026/05/14 – 2026/05/14

Dissertation Title: Efficient and Generalizable Robot Learning via Physical, Geometric, and Semantic Priors

Speaker: Jianshu Hu, Ph.D. candidate at SJTU Global College

Time: May 14 from 10:00-11:00 a.m., 2026 (Beijing Time)

Location: Room 503, Longbin Building

Abstract

The pursuit of general-purpose robotic agents, which are capable of performing diverse tasks in unstructured environments, stands as one of the grand challenges of modern robotics domain. While data-driven methods such as Reinforcement Learning (RL) and Imitation Learning (IL) have revolutionized the field, enabling robots to master complex skills from pixel inputs, they remain hampered by two critical bottlenecks: sample inefficiency (the need for a large amount of interactions to learn a task) and limited generalization ability (failure when facing novel tasks or environments). This thesis addresses these limitations by developing more sample-efficient and generalizable methods that systematically inject additional knowledge into the learning process.

Consider the mechanisms by which humans acquire new manipulation skills. First, humans learn to manipulate single object by learning environmental dynamics through physical interaction, refining its movements based on the consequences of executed actions. Second, humans efficiently leverage geometric priors, such as symmetry, to generalize across diverse objects and scenarios. Finally, humans utilize existing semantic knowledge to identify objects, interpret scene relationships and plan complex tasks. Inspired by these biological capabilities, this thesis proposes and investigates three components for building efficient and generalizable robot learning algorithms: (1) Physical Priors: The effective integration of learned dynamics models to accelerate policy learning and improve generalization. (2) Geometric Priors: The systematic incorporation of geometric prior by applying data augmentation in robotic tasks. (3) Semantic Priors: The transfer of knowledge from pre-trained models to enhance the perceptual and planning capabilities of the policy.

Biography

Jianshu Hu is a final-year Ph.D. candidate at SJTU Global College in Shanghai. He is currently advised by Prof. Yutong Ban and Prof. Paul Weng. He is interested in robot learning and robot manipulation. His current research focuses on improving sample-efficiency and generalization ability of robot learning algorithms by exploiting data augmentation, leveraging pre-trained models, and learning a dynamics model/world model.