Workshop on effective Representations, Abstractions, and Priors for Robot Learning (RAP4Robots)
Monday, 29th May @ ICRA 2023 @ CAPITAL SUITE 8
Description
This workshop aims to facilitate discussion about key issues related to scalable and generalizable robot learning, with a focus on three main axes: state/action representations, abstractions, and behavior priors. While robot learning holds the promise of endowing robots with complex skills, in practice, the scalability and generalization of skills are still an open problem. One pressing issue is the type and quality of representations that we could learn, for e.g., when learning in simulation or when learning in a different domain. Moreover, the sample-efficient learning of modular skills is crucial for robotics, and in this regard, the use of priors, either classical or learned, can greatly benefit learning. This makes researchers think about the best way to induce prior knowledge into a learning system, but the representation of such priors should “fit” the follow-up tasks. Moreover, the decomposition of complex long-horizon tasks into a sequence of simpler ones is a well-known strategy in robot planning. However, the type and levels of abstractions needed are open questions, and their representation and connection to observations are still challenging.
Workshop panels will discuss fundamental open research questions, such as:
How can we effectively represent multisensorial data and experiences across different instances of the robot’s operating life?
How can we take advantage of robotic priors, scene structure, and demonstrations to accelerate robotic learning?
What and how many levels of abstractions should we consider for long-horizon robotic tasks?
How can we achieve effective lifelong robot learning, while avoiding catastrophic forgetting?
This workshop will bring together a cohort of top researchers, both early-career and established ones, to discuss these highly relevant problems, and open up new directions for developing systems and algorithms for robot learning that respect top-down and bottom-up compatibility (in terms of representations, abstractions, and sequencing strategies). After the workshop, we plan to organize a special issue at a major robotics journal (e.g., IEEE RA-L) on the topic of representations, abstractions, and priors for robot learning, also driven by the outcomes of the workshop.
Topics of interest:
· Representation learning
· Multisensory learning and fusion
· Representation alignment for Human-Robot Interaction
· Structured priors
· Model-based RL and world models
· Hierarchical learning and planning
· Skill composition and decomposition
· Transfer learning
Accepted Papers
Spotlights:
A Lagrangian Inspired Polynomial Kernel for Robot Dynamics Identification
LIBERO: Benchmarking Knowledge Transfer in Lifelong Robot Learning
Learning to Read Braille: Bridging the Tactile Reality Gap with Diffusion Models
Human-In-The-Loop Task and Motion Planning for Imitation Learning
Effective State Representation for Learning Hierarchical Robust Bipedal Locomotion
CLIP-Fields: Weakly Supervised Semantic Fields for Robotic Memory
VL-Fields: Towards Language-Grounded Neural Implicit Spatial Representations
Neural SE(3)-equivariant Representations for Object-based SLAM
Detect-O-Con: Detecting Object Co-occurrences via NeRF representations
Posters:
Bayesian Optimization Enhanced by Transfer Learning for 2D Camera-Based Robotic Manipulation
Enhancing Robot Learning through Learned Human-Attention Feature Maps
Unified Representation for Learning Object-Centric Articulation Models Using Geometric Algebra
Real-Time Reinforcement Learning for Vision-Based Robotics Utilizing Local and Remote Computers
Touch Primitives for Gripper-Independent Haptic Object Modeling
Investigating the Impact of Action Representations in Policy Gradient Algorithms
Learning Generalizable Pivoting Skills with Object Feature Based State/Action Projections
Efficient Automatic Tuning for Data-driven Model Predictive Control via Meta-Learning
Motion Tasks Representation: Extracting Knowledge from Human Experts
Speakers
University of Toronto/ NVIDIA, Canada
Stanford University, USA
Imperial College London, UK
Facebook AI (FAIR), USA
Deepmind, UK
University of the Witwatersrand, South Africa
Brown University, USA
Program
Organizers
TU Darmstadt, Germany
Stanford University, USA
University of Tokyo, Japan
Carnegie Mellon University, USA
University of Sydney/ NVIDIA, Australia
TU Darmstadt, Germany
TU Darmstadt, Germany
Carnegie Mellon University, USA
Extended program committee:
Alap Kshirsagar
An Thai Le
Sabin Grube Doiz
Daniel Palenicek
Firas Al-Hafez
Sarvesh B Patil
Gabriele Tiboni
Joao Carvalho
Saumya Saxena
Junning Huang
Kay Hansel
Sohan Rudra
Kevin Zhang
Mark Lee
Sophie Lueth
Patrick Callaghan
Rickmer Krohn
Xiaolin Lin
Sponsors:
Supported by the IEEE RAS Technical Committees on:
Robot Learning
Cognitive Robotics