IPAB Workshop - 7/5/26 Speaker: Matias Mattamala Aravena Title: Robot Representations for Real-World Autonomy Abstract: This is my first IPAB presentation since I joined Informatics, so I'll use the opportunity to introduce part of my past and current research in perception and scene representations for robot autonomy. Motivated by diverse real-world applications of robot monitoring, exploration, and assistance, I will show different examples of how building or choosing the right representation (e.g., topological map, local maps, embeddings) enables the implementation of effective perception-action loops to make robots that operate autonomously with onboard sensing and compute. To conclude, I will share some of my current work extending these principles to assistive robots in daily activities, as part of the Centre for AI in Assistive Autonomy. Speaker: Andreas Sochoupoulos Title: From imitation learning to RL fine-tuning for robots that learn from experience Abstract: Diffusion and flow matching policies have recently demonstrated remarkable performance in robotic applications by accurately capturing multimodal robot trajectory distributions, especially in the context of Vision Language Action (VLA) models. However, high quality policy performance also requires high quality demonstrations which are often hard to get. In this talk I will present our work tackling the problem of fine-tuning suboptimal diffusion/flow based policies using the robot's experience. Robots should be able to self-improve through their own successes and failures. For this reason, we introduce a new methodology for finetuning flow policies and flow-based VLAs using either offline or online RL, while at the same time reducing the computational complexity of action generation. May 07 2026 13.00 - 14.00 IPAB Workshop - 7/5/26 Matias Mattamala Aravena & Andreas Sochoupoulos MF2
IPAB Workshop - 7/5/26 Speaker: Matias Mattamala Aravena Title: Robot Representations for Real-World Autonomy Abstract: This is my first IPAB presentation since I joined Informatics, so I'll use the opportunity to introduce part of my past and current research in perception and scene representations for robot autonomy. Motivated by diverse real-world applications of robot monitoring, exploration, and assistance, I will show different examples of how building or choosing the right representation (e.g., topological map, local maps, embeddings) enables the implementation of effective perception-action loops to make robots that operate autonomously with onboard sensing and compute. To conclude, I will share some of my current work extending these principles to assistive robots in daily activities, as part of the Centre for AI in Assistive Autonomy. Speaker: Andreas Sochoupoulos Title: From imitation learning to RL fine-tuning for robots that learn from experience Abstract: Diffusion and flow matching policies have recently demonstrated remarkable performance in robotic applications by accurately capturing multimodal robot trajectory distributions, especially in the context of Vision Language Action (VLA) models. However, high quality policy performance also requires high quality demonstrations which are often hard to get. In this talk I will present our work tackling the problem of fine-tuning suboptimal diffusion/flow based policies using the robot's experience. Robots should be able to self-improve through their own successes and failures. For this reason, we introduce a new methodology for finetuning flow policies and flow-based VLAs using either offline or online RL, while at the same time reducing the computational complexity of action generation. May 07 2026 13.00 - 14.00 IPAB Workshop - 7/5/26 Matias Mattamala Aravena & Andreas Sochoupoulos MF2
