A hierarchical learning framework using large language models and semantic keypoints enables generalizable clothes manipulation across diverse clothes categories and tasks.
Plants can be leveraged as actuators with integrated power sources to create biodegradable robots capable of locomotion and object manipulation.
달 표면 자율 탐사를 위해 다양한 센서 데이터와 고정밀 지면 진실 레이블을 제공하는 LuSNAR 데이터셋
LuSNARデータセットは、月面探査のための自律型環境認識とナビゲーションを包括的に評価するための、マルチタスク、マルチシーン、マルチラベルのベンチマークデータセットである。
The LuSNAR dataset provides a multi-task, multi-scene, and multi-label benchmark for evaluating autonomous perception, navigation, and reconstruction algorithms for lunar exploration.
A novel system that enables rigid robots to learn dexterous, contact-rich manipulation tasks from a few demonstrations, incorporating a teleoperation interface with haptic feedback and a method called Comp-ACT that learns variable compliance control.
A unified vision-based whole-body-control parkour policy enables a humanoid robot to autonomously overcome various challenging obstacles, including jumping on platforms, leaping over hurdles, and traversing different terrains.
Incorporating the inductive bias of action locality into the design of a visuomotor policy framework can significantly boost sample efficiency in robotic manipulation tasks.
개방형 세계에서 로봇이 이동할 때 주변 로봇들로부터 지식을 전달받아 자신의 위치를 추정하는 새로운 학습 방식을 제안한다.
This work introduces a novel data-free recursive distillation scheme for open-world distributed robot systems, where a student robot can ask various types of teacher robots, including uncooperative, untrainable, or black-box models, for guidance in unfamiliar workspaces.