The field of robotics is witnessing significant advancements in robot learning and task planning, driven by innovations in large language models, vision-language-action models, and imitation learning. Researchers are exploring new approaches to improve the efficiency, adaptability, and generalization of robotic systems, enabling them to perform complex tasks in diverse environments. Notable developments include the integration of large language models with planning domain definition language, the use of geometric fabrics for safe and stable imitation learning, and the introduction of novel frameworks for continual adaptation, visual data augmentation, and cross-task invariance. These advancements have the potential to enhance the capabilities of robots in various applications, from household tasks to industrial manufacturing. Some noteworthy papers in this regard include LLM+MAP, which introduces a bimanual planning framework that integrates LLM reasoning and multi-agent planning, and TamedPUMA, which proposes an IL algorithm augmented with geometric fabrics for safe and stable imitation learning. Additionally, papers like Gemini Robotics and MoLe-VLA demonstrate significant improvements in robot learning and task planning, with Gemini Robotics introducing a generalist Vision-Language-Action model and MoLe-VLA proposing a dynamic layer-skipping vision language action model.
Advances in Robot Learning and Task Planning
Sources
LLM+MAP: Bimanual Robot Task Planning using Large Language Models and Planning Domain Definition Language
TamedPUMA: safe and stable imitation learning with geometric fabrics
Efficient Continual Adaptation of Pretrained Robotic Policy with Online Meta-Learned Adapters
RoboEngine: Plug-and-Play Robot Data Augmentation with Semantic Robot Segmentation and Background Generation
AdaWorld: Learning Adaptable World Models with Latent Actions
CubeRobot: Grounding Language in Rubik's Cube Manipulation via Vision-Language Model
DataPlatter: Boosting Robotic Manipulation Generalization with Minimal Costly Data
Dita: Scaling Diffusion Transformer for Generalist Vision-Language-Action Policy
Gemini Robotics: Bringing AI into the Physical World
MoLe-VLA: Dynamic Layer-skipping Vision Language Action Model via Mixture-of-Layers for Efficient Robot Manipulation
Offline Action-Free Learning of Ex-BMDPs by Comparing Diverse Datasets
OminiAdapt: Learning Cross-Task Invariance for Robust and Environment-Aware Robotic Manipulation
Cultivating Game Sense for Yourself: Making VLMs Gaming Experts
Cooking Task Planning using LLM and Verified by Graph Network
Data-Agnostic Robotic Long-Horizon Manipulation with Vision-Language-Guided Closed-Loop Feedback
CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models
REMAC: Self-Reflective and Self-Evolving Multi-Agent Collaboration for Long-Horizon Robot Manipulation
Robust Offline Imitation Learning Through State-level Trajectory Stitching
ActionStudio: A Lightweight Framework for Data and Training of Action Models