Teaching robots complicated manipulation skills through observation of human demonstrations has shown promising results. Providing extensive manipulation demonstrations is time-consuming and labor costly, making it challenging to scale up this paradigm to real-world long-horizon operations. However, not all facets of a task are created equal. A new study by NVIDIA and Georgia Institute of Technology…

In many domains that involve machine learning, a widely successful paradigm for learning task-specific models is to first pre-train a general-purpose model from an existing diverse prior dataset and then adapt the model with a small addition of task-specific data. This paradigm is attractive to real-world robot learning since collecting data on a robot is…

Training robots to perform various manipulation behaviors has been made possible by imitation learning from human demonstrations. One popular method involves having human operators teleoperate with robot arms through various control interfaces, producing multiple demonstrations of robots performing different manipulation tasks, and then using the data to train the robots to perform these tasks independently.…

In robotics, researchers face challenges in using reinforcement learning (RL) to teach robots new skills, as these skills can be sensitive to changes in the environment and robot structure. Current methods need help generalizing to new combinations of robots and tasks and handling complex, real-world tasks due to architectural complexity and strong regularisation. To tackle…

A team of researchers from MIT and the Institute of AI and Fundamental Interactions (IAIFI) has introduced a groundbreaking framework for robotic manipulation, addressing the challenge of enabling robots to understand and manipulate objects in unpredictable and cluttered environments. The problem at hand is the need for robots to have a detailed understanding of 3D…

A team of researchers from the University of Illinois Urbana-Champaign, Carnegie Mellon University, Georgia Institute of Technology, University of California Berkeley, Meta AI Research, and Mistral AI has developed a universal navigation system called GO To Any Thing (GOAT). This system is designed for extended autonomous operation in home and warehouse environments. GOAT is a…

Researchers from MIT and Meta AI have developed an object reorientation controller that can utilize a single depth camera to reorient diverse shapes of objects in real-time. The challenge addressed by this development is the need for a versatile and efficient object manipulation system that can generalize to new conditions without requiring a consistent pose…