Nimble
Nimble is training robots to learn human skills and adapt to perform unseen, novel skills!
Background
Nimble is training robots to learn human skills and adapt to perform unseen, novel skills. We utilize simulated and real-world data to teach an articulated robot arm a set of human skills such as grasping objects, using tools, or drawing. With different machine learning techniques, we can also teach it to perform tasks it has never learned before!
Goals
Adaptation to Unseen Tasks - use techniques to utilize information from training data to unseen and novel tasks or environments.
Long-Sequence Planning: maintain quality of performance across sequences of several tasks, develop planning skills to smoothly execute long, difficult sequences.
Relevant Skills
Members of the Nimble Team primarily work on software tasks for the project. This includes fine-tuning of the vision-language-action model, building a pipeline for converting model inference to movement outputs, and designing simulation tasks for training our robotic arm, among other tasks. There are also several design-based tasks, including real-world testing environment construction and gripper/finger prototyping, as well as real-world experimentation and data collection. All members also stay up-to-date with relevant research and algorithms and come up with ways to apply new research to our project’s goals.