LICARSA 2023 Abstracts

Area 1 - Learning in Control and Robotic Systems Applications

Full Papers
Paper Nr: 5

Improving Reward Estimation in Goal-Conditioned Imitation Learning with Counterfactual Data and Structural Causal Models


Mohamed K. Jabri, Panagiotis Papadakis, Ehsan Abbasnejad, Gilles Coppin and Javen Shi

Abstract: Imitation learning has emerged as a pragmatic alternative to reinforcement learning for teaching agents to execute specific tasks, mitigating the complexity associated with reward engineering. However, the deployment of imitation learning in real-world scenarios is hampered by numerous challenges. Often, the scarcity and expense of demonstration data hinder the effectiveness of imitation learning algorithms. In this paper, we present a novel approach to enhance the sample efficiency of goal-conditioned imitation learning. Leveraging the principles of causality, we harness structural causal models as a formalism to generate counterfactual data. These counterfactual instances are used as additional training data, effectively improving the learning process. By incorporating causal insights, our method demonstrates its ability to improve imitation learning efficiency by capitalizing on generated counterfactual data. Through experiments on simulated robotic manipulation tasks, such as pushing, moving, and sliding objects, we showcase how our approach allows for the learning of better reward functions resulting in improved performance with a limited number of demonstrations, paving the way for a more practical and effective implementation of imitation learning in real-world scenarios.

Nr: 198

Long-Term Dynamical Predictions for Autonomous Underwater Vehicles


Pierre Nicolay, Mykhaylo Marfeychuk, Yvan Petillot and Ignacio Carlucho

Abstract: In this work, we present a model structure and a novel cost function to learn dynamical models that can be used for stable long-term predictions. We use the topological and gradient properties of the SE3 Lie-Group on which Autonomous Underwater Vehicles (AUVs) operate.

Nr: 199

Simulating Robot Navigation in Outdoor Unstructured Environments


Giuseppe Vecchio, Simone Palazzo, Dario C. Guastella, Riccardo Emanuele Sarpietro, Ignacio Carlucho, Stefano Albrecht, Giovanni Muscato and Concetto Spampinato

Abstract: Autonomous ground robot navigation in outdoor unstructured environments is challenging due to the complexity of real-world scenes. Gathering data for model training in these conditions is costly and risky, emphasizing the need for simulation environments. However, there is a lack of simulators for such environments. To fill this gap, we introduce MIDGARD, an open-source platform for outdoor navigation based on Unreal Engine, offering photorealistic environments, procedural scene generation and compatibility with OpenAI Gym and overcoming the limitations of simulation platforms like CARLA, AirSim, HABITAT and OAISYS. MIDGARD is a flexible, high-performance, open-source platform designed for outdoor navigation in unstructured environments. It allows users to configure and extend the core engine and is equipped with a wide suite of ready-to-use sensors. It also provides access to internal state variables for training or reward computation in a reinforcement learning setting. One of MIDGARD's key features is its ability to generate varying and dynamic simulated scenes on the fly. The scene generation process in MIDGARD is fully procedural and requires no human intervention beyond an initial setup stage. Scene descriptors define the base map for the scene type and a set of placeable world objects, each defined by its 3D polygonal mesh, attributes and instantiation constraints. The virtual agent in the simulated scene is defined by two component modules: the perception module and the control suite. The perception module provides a full set of sensors designed for navigation, including vision sensors and low-level sensors for agent state measurements. The control suite includes two types of 4-wheel vehicles.

Nr: 200

Vision-Based Pose Estimation of Waste Objects for Robotic Grasping


Francesco Cancelliere, Luca Reitano, Dario C. Guastella, Giuseppe Sutera and Giovanni Muscato

Abstract: In this work we present a comparison of vision-based pose estimation methods for the robotic grasping of waste objects. We assume to detect the objects with an RGBD camera vertically pointing towards the ground. Unlike other methods requiring long-trained reinforcement learning agents or point cloud data, we propose simple yet effective methods essentially relying on 2D object detection.