Modular and Lifelong Learning for Developing General Purpose Robotic Agents

对于通才机器人来说，强化学习或从交互中学习是一条很有前途的途径，它可以通过额外的训练来学习和提高它们的表现。本研究计划的长期目标（10-15年）是通过(1)学习算法来实现在现实世界中的有效训练，以及(2)开发有原则的通用奖励函数来鼓励机器人学习与人类相似的各种行为，从而解决在将交互方法学习应用于真实机器人代理时的两个限制。本研究以不受模拟过度影响的解决方案为指导，旨在针对现实世界中真实机器人的工作方法，以避免潜在的偏差。这些现实世界的假设为学习算法引入了额外的约束。然而，这些约束催生了一个更独立的学习系统，具有更像人类的学习能力。该研究计划有三个关键的短期（5年）目标，这些目标是限制（RL）范式产生具有人类水平技能的代理的障碍。目标1：通过学习模块化和可重用的决策过程，构建算法，实现跨mdp和形态学的泛化。目标2：开发持续学习系统，使智能体能够在现实世界中独立训练。目标3：创建一般和原则性的奖励函数，使代理能够学习对未来任务有用的各种技能。

Reinforcement learning or learning from interaction is a promising route to generalists robots that can learn and improve their performance with additional training. The long-term objective (10-15 years) of this research program is to address two limitations in applying learning from interaction methods on real robotic agents via (1) learning algorithms to enable efficient training in the real world and (2) develop principled general purpose reward functions to encourage robots to learn diverse behaviours similar to humans. This research is directed by solutions that are not overly influenced by simulation and aims to target methods that work on real robots in the real world to avoid potential bias. These real-world assumptions introduce additional constraints on the learning algorithms. However, these constraints induce a more independent learning system with more human-like learning capabilities. This research program has three key short-term (5 year) objectives that are hurdles limiting the (RL) paradigm from producing agents with human-level skill. Aim 1 Build algorithms that enable generalization across MDPs and morphologies by learning modular and reusable decision making processes. Aim 2: Develop continual learning systems to enable agents that can functionally and independently train in the real-world. Aim 3: Create general and principled reward functions that enable agents to learn diverse skills that are useful for future tasks.