Advances in Robot Learning and Control

The field of robot learning and control is moving towards more generalizable and flexible approaches. Researchers are exploring methods that enable robots to learn from diverse demonstrations, adapt to new situations, and perform complex tasks. One notable direction is the use of large language models (LLMs) as numerical optimizers for robot self-improvement, allowing for iterative learning and adaptation of robot behavior. Another significant trend is the development of hierarchical robot planning frameworks that can integrate kinematic skills and closed-loop motor controllers, enabling the use of diverse pre-learned skills in hierarchical robot planning. Noteworthy papers in this area include CIVIL, which proposes a causal and intuitive visual imitation learning approach that enables robots to learn from humans by indicating task-relevant features, and DeCo, which presents a task decomposition and skill composition framework for zero-shot generalization in long-horizon 3D manipulation tasks. LLM-iTeach is also notable, as it introduces a novel interactive imitation learning framework that utilizes an LLM as an interactive teacher to enhance agent performance. LangWBC and SAS-Prompt are also worth mentioning, as they demonstrate the potential of LLMs in whole-body control and numerical optimization for robot self-improvement.

Sources

Beyond Task and Motion Planning: Hierarchical Robot Planning with General-Purpose Policies

CIVIL: Causal and Intuitive Visual Imitation Learning

Offline Learning of Controllable Diverse Behaviors

Instrumentation for Better Demonstrations: A Case Study

Generalization Capability for Imitation Learning

SAS-Prompt: Large Language Models as Numerical Optimizers for Robot Self-Improvement

Leveraging Pre-trained Large Language Models with Refined Prompting for Online Task and Motion Planning

LangWBC: Language-directed Humanoid Whole-Body Control via End-to-end Learning

LLM-based Interactive Imitation Learning for Robotic Manipulation

DeCo: Task Decomposition and Skill Composition for Zero-Shot Generalization in Long-Horizon 3D Manipulation

Built with on top of