IPAB Workshop-25/07/2024

Speaker: Hefan Wang

Title: Hybrid Loco-Manipulation in Quadruped Robots: Leveraging Optimization and Learning-Based Strategies

Abstract: Loco-manipulation in quadruped robots requires complex coordination of limb movements and body stability, especially in dynamic and unpredictable environments. Traditional optimisation and learning-based methods each show potential but have limitations when used alone. Optimisation methods excel in precise motion control and long-term trajectory planning but lack adaptability and require long convergence times. Conversely, reinforcement learning (RL) offers adaptability and robustness but struggles with long-horizon planning due to sparse reward signals. This work proposes an innovative hybrid framework that combines these approaches, utilising optimisation for global planning and dynamic constraint adherence, and RL for adaptable, robust movement control.

Speaker: Kale-Ab Tessera

Title: Enabling Specialisation in Multi-Agent RL

Abstract: Multi-Agent Reinforcement Learning (MARL) is crucial for advancing complex, real-world applications where multiple agents must learn to cooperate or compete, leading to breakthroughs in fields such as autonomous systems, robotics, and strategic games.

Parameter sharing is a key ingredient for scalable MARL algorithms. However, it has been shown to limit the performance of MARL algorithms in certain settings where the optimal policy requires individual agents to behave differently. We propose measures of agent diversity and validate that, in certain settings, PPO with parameter sharing fails to learn diverse enough policies to solve problems, whereas agents with non-shared policies can. We also show that common parameter-sharing practices, such as conditioning on agent IDs, are often insufficient for enabling these shared policies to learn agent-specific behaviours. Finally, we discuss some of the challenges of training these systems and propose methods to enable specialization in these settings.

Speaker: Haocheng Yuan

Title: Accessible Computer-Aided Design

Abstract: Computer-Aided Design (CAD) is the industry standard for representing 3D shapes as sequences of geometric instructions, known as CAD programs. These programs are compact, expressive, and allow for parametric editing of shapes. However, designing and editing CAD shapes require specific training in particular software. To make CAD more accessible, we developed CADTalk, the first algorithm for semantic commenting of CAD programs, and DiffCSG, the first differentiable rasterizer for CAD primitives with Boolean operators. This presentation will introduce these two methods, discuss their contributions to accessible CAD, and explore future directions in this field.