Selected Projects
Diffusion Model Fine-Tuning With Constrained Optimization
Diffusion models have been successful in a number of fields in recent years due to their ability to obtain a synthetic probability distribution for a given dataset. How- ever, they are likely trained on general data and can generate undesired results for specific tasks and use cases. In this work, we propose a simple pipeline to fine- tune unconditional diffusion models via a constrained optimization process. We show that this formulation has broad compatibilities with straight-forward score functions in diverse domains of applications by experiments in (i) removing the class of digits in an MNIST data set, (ii) simulating safety constraints for trajec- tory planning, and (iii) optimizing with pairs of expert preferences on polymer generation tasks. We show that our framework is robust and easy to implement, only requiring an additional coarse penalty term for all of the experiments we demonstrate.
MatchingPennies(2nd-place one-shot track negotiation agent at SCML 2024)
The MatchingPennies Agent employs concurrent negotiation and a heuristic function to identify the most profitable subsets of current offers in a setting of multi-agent negotiation provided by negmas. The ability to balance between fulfilling the desired quantities and negoatiating for the best price gives its name. The agent won the 2nd place in the SCML2024 one-shot track.
Pipelining Natural Language to Multi-Robot Task Allocation
This initiative aims to simplify commanding a team of robots through a logical intermediary. By translating natural language commands into linear temporal logic(LTL), we've enhanced the success rate of task execution. The approach supports complex commands and collaborative tasks between human and teams of robots.