USC researchers will present three papers at the 2026 Robotics: Science and Systems (RSS) conference, with research spanning robotic foundation models, humanoid loco-manipulation, robotic reward models and robot manipulation learning.
Held July 13–17 in Sydney, Australia, RSS 2026 is one of the world’s largest robotics conferences, bringing together researchers from academia and industry to share the latest advances in robotics through peer-reviewed papers, workshops, tutorials and technical presentations.
Faculty and students from the USC Viterbi School of Engineering‘s Thomas Lord Department of Computer Science and the Ming Hsieh Department of Electrical and Computer Engineering, and USC Mark and Mary Stevens School of Computing and AI, will represent the university throughout the conference.
In addition to presenting research papers, USC researchers are contributing as workshop organizers, invited speakers and area chairs, reflecting the university’s continued leadership across the robotics community.
This year, computer science and electrical and computer engineering assistant professor Erdem Bıyık will serve as one of RSS’s area chairs, while computer science assistant professor Yue Wang will lead and speak at two workshops. Wang is one of the organizers of the workshop, “Post-Training for Robotics Foundation Models: From Pretrained Policies to Real-World Mastery,” and will speak at the “Beyond Rigidity: Deformable and Articulated Robotic Object Manipulation” workshop on the conference’s opening day.
This year’s has an acceptance rate of approximately 32%.
RSS also shares a long-standing connection with USC. Gaurav S. Sukhatme, USC Viterbi’s Dean (Interim) and director of the USC Mark and Mary Stevens School of Computing and AI, co-founded the conference in 2004, helped establish the Robotics: Science and Systems Foundation and served as program chair for its inaugural conference in 2005. USC has since played a central role in the conference’s growth, hosting RSS in 2011 and again in 2025, when the university welcomed the largest edition in the conference’s history with nearly 1,300 attendees.
USC @ RSS 2026 Research Spotlights #
Helping Human-Like Robots Master Everyday Chores
To become truly useful in homes, hospitals and workplaces, humanoid robots must be able to perform everyday tasks that people often take for granted, from making coffee and wiping tables to pushing carts. In the paper, “Ψ0: An Open Foundation Model Towards Universal Humanoid Loco-Manipulation,” USC researchers introduce a new foundation model that helps robots learn these complex skills more efficiently.
Rather than relying solely on costly robot training data, the team developed Ψ0 (Psi-Zero), a two-stage training framework that first learns general movement patterns from more than 800 hours of human video before refining those skills using just 30 hours of robot-specific data. Because robots move differently than humans, directly copying human demonstrations has long been a challenge. By bridging that gap, Ψ0 enables humanoid robots to complete long, multi-step tasks more effectively than previous approaches, bringing researchers closer to versatile robots capable of assisting people with a wide range of real-world activities.
Teaching Robots to Learn From Their Mistakes
People often learn new skills by making mistakes, but robots are typically trained using only perfect examples. In the paper “Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons,” USC researchers introduce a new reward model that helps robots learn not only from successful attempts but also from failures.
Instead of relying exclusively on expert demonstrations, Robometer compares different attempts at completing a task to evaluate which actions move a robot closer to success. Trained on more than 1 million video clips of both humans and robots, the system enables robots to better judge their own progress, automatically recognize when they have made mistakes and improve through repeated practice. The approach could help robots learn everyday manipulation tasks, such as placing a bowl onto a plate, more efficiently while becoming more adaptable in real-world environments.
Giving Robots 3D Vision for Precise Hand-Eye Coordination
Many robots rely on conventional 2D camera images to understand their surroundings, but those images often lack the depth information needed for precise manipulation tasks. In the paper “CLAMP: Contrastive Learning for 3D Multi-View Action-Conditioned Robotic Manipulation Pretraining,” USC-affiliated researchers introduce a new training framework that helps robots better understand the three-dimensional world around them.
CLAMP combines 3D data from multiple viewpoints, including cameras mounted on a robot’s wrist, with the robot’s actions and language instructions to build a richer understanding of how objects are positioned in space. This enhanced hand-eye coordination allows robots to learn high-precision manipulation tasks, such as inserting a pen into a narrow container or opening a specific drawer, more quickly and accurately in both simulated and real-world environments.
USC-Affiliated Papers #
(USC authors bolded) Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons
Anthony Liang, Yigit Korkmaz, Jiahui Zhang, Minyoung Hwang, Abrar Anwar, Sidhant Kaushik, Aditya Shah, Alex S. Huang, Luke Zettlemoyer, Dieter Fox, Yu Xiang, Anqi Li, Andreea Bobu, Abhishek Gupta, Stephen Tu, Erdem Bıyık, Jesse Zhang
Ψ0: An Open Foundation Model Towards Universal Humanoid Loco-Manipulation
Songlin Wei, Hongyi Jing, Boqian Li, Zhenyu Zhao, Jiageng Mao, Zhenhao Ni, Sicheng He, Jie Liu, Xiawei Liu, Kaidi Kang, Sheng Zang, Weiduo Yuan, Marco Pavone, Di Huang, Yue Wang
CLAMP: Contrastive Learning for 3D Multi-View Action-Conditioned Robotic Manipulation Pretraining
I-Chun Arthur Liu, Krzysztof Choromanski, Sandy Huang, Connor Schenck
Published on July 15th, 2026
Last updated on July 15th, 2026