Generalizable Robot Learning from In-the-Wild Datasets
Abstract
Generalizable Robot Learning from In-the-Wild DatasetsChelsea Finn, Stanford University.Program Officer: Marc SteinbergApproved forPublic ReleaseAbstract. Robots that are autonomous and versatile have the potential to be tremendously usefulfor a range of manual,tedious, or dangerous tasks that are currently performed entirely by humans.However, in order to be successful in performing manipulation tasks in diverse, dynamic, and unstructuredenvironments, robots must be able to learn continuously#across tasks, environments,and even experimental set-ups#perpetually accumulating their experiences and learning from them.This setting, however, presents massive challenges for current robot learning approaches, whichlargely assume carefully-designed experimental set-ups and state representations, narrow single-tasklearning problems, and supervision in the form of reward feedback or labels. In this project,we will develop algorithms for training dynamics models, behavioral priors, and goal representationsfrom scalable sources of data, including heterogeneous data from previous experiments andinternet-scale datasets of human behavior.While preliminary results on this project have shown promising signs of generalization byusing multi-environment data, the requested equipment will substantially enhance this research byenabling the models to scale to much broader datasets, such as datasets with hundreds of tasksand environments rather thanonly a handful of tasks and environments. Specifically, this proposalbudgets for three GPU servers with state-of-the-art GPUs that allow for training large models thatcan fit broad datasets. The largest server contains 8 GPUs with 80 GB of memory, which makes itpossible to train models that are more than 6x larger than our largest previous models, and morethan 10x larger when incorporating memory optimization techniques. This equipment will henceallow us to scale to 10x broader data, and consequently study how robots can generalize behaviorsto entirely new environments. This ability to operate successfully in new environments is essentialfor robots tobe useful in real, unstructured environments.Technical Approach. The proposed research will build and study data-driven algorithms forthree problem areas that are central to robot learning.Part 1: Visual World Models. We will study how to train deep predictive models of the robot#ssensory inputs with diverse, cross-domain data, including data of humans, and how to use thesemodels for performing a breadth of manipulation tasks in the real world.Part 2: Meta-Learning. We still study how meta-learning can be used to learn priors over behaviorto enable fast adaptation to new goals and domains, including composing skills into new behaviors.Part 3: Task Inference. We will study how to learn unified task representations from diverse,multimodal data sources that make it possible for robots to acquire sets of tasks to learn fromoffline data and to represent goals from human-provided task specifications.Anticipated Outcomes. The proposed research will advance the state-of-the-art of robot learning,enabling robots to generalize much more broadly and to perform tasks of greater complexityby learning powerful and generalizable representations from diverse data sources.Impact on DoDCapabilities. Robotic systems that can continuously learn from broad datasetswill be enable robots to have impact around autonomously performing a variety of useful tasks inunstructured environments, including cleaning, maintenance, repair, and construction.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 11, 2023
- Source ID
- N000142312845
Entities
People
- Chelsea Finn
Organizations
- Office of Naval Research
- Stanford University
- United States Navy