Grants and Contracts per year
Grants and Contracts Details
Robot-led expletory missions in unknown and harsh environments, such as Mars and Moon need robots to make planning decisions for conducting missions. Given that such environments are unfriendly environments for wireless communication channels, relying heavily on the ground station to make such planning decisions is not tenable. Thus, robots must demonstrate some semblance of autonomy. Planning could benefit from the ability to predict changes to the communication channel if a robot takes a given movement action. With such an ability, robots could plan movements to maintain a strong communication signal. Using centralized deep learning for such a prediction can be costly due to communicating large raw data to the ground station. Such communication could incur a great deal of communication cost. In this work, we push model fitting/training to the robots and use federated learning (FL) to aggregate model weights fitted by multiple different robots. Recent work has shown FL can reduce communication cost due to no raw data communication. However, there exists a trade- off between the accuracy obtained by a certain frequency of aggregation for robot-trained models and the overall communication cost. Increasing aggregation could increase accuracy while incurring additional communication cost. This proposed work will study this trade-off to support simultaneous planning and learning missions in such unknown environments (e.g., lunar and Martian caves) for robot-led missions.
|Effective start/end date
|1/1/23 → 5/31/24
- National Aeronautics and Space Administration
Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.