Collaborative Learning and Adaptive Robots (CLeAR) Lab.
Below, you’ll find our contributions to robot/machine learning. Much of our work is on learning from (or interacting with) humans.
We show that a single unconditional diffusion model performs competitively in out-of-distribution detection tasks by measuring the rate-of-change and curvature of diffusion paths connecting data samples to the standard normal distribution.
We develop a safe planning method for trajectory generation by sampling from diffusion model under different LTLf constraints.
We extend gradient flow methods to a variety of high-quality image synthesis tasks using a novel density ratio learning method.
We introduce PhysiCLeaR, an annotated dataset of everyday objects and tactile readings collected from a Gelsight Mini sensor, as well as Octopi, a system that leverages both tactile representation learning and large vision-language models to perform physical reasoning and inference, given tactile videos of multiple objects.
We develop interpolant policies that leverage informative source distributions for imitation learning.
We propose an optimization-based grasp synthesis framework, GRaCE, to generate context-specific grasps in complex scenarios. We test GRaCE in a simulator and a real-world grasping tasks.
We introduce a novel framework centered around the Probable Object Location (POLo) score, which allows the agent to make data-driven decisions for efficient object search.
We apply techniques from continual learning to the problem of selective forgetting in deep generative models. Our method, dubbed Selective Amnesia, allows users to remap undesired concepts to user-defined ones.
We study a model of multi-population learning with heterogenous beliefs.
We construct a decomposed latent state space model for perspective-taking for human robot interaction.
We explore the potential of LLMs to act as zero-shot human models for HRI. We contribute an empirical study and case studies on a simulated table-clearing task and a new robot utensil-passing experiment.
We propose a sampling-based grasp synthesis framework, GraspFlow, to generate context-specific grasps. We test GraspFlow in a real-world table-top grasping task.
We develop a family of stable continuous-time neural state space-models.
We contribute an empirical study into the effectiveness of LLMs, specifically GPT-3.5 variants, for the task of natural language goal translation to PDDL.
Transfer source policies to a target reinforcement learning task with safety constraints using Successor Features.
We examine the problem of observed adversaries for deep policies, where observations of other agents can hamper robot performance.
This paper proposes SCALES, a general framework that translates well-established fairness principles into a common representation based on CMDPs.
We develop an accurate physics-inspired model for describing how a population of Q-learning agents adapt as they interact.
Inspired by Social Projection Theory, we use the robot's self model to efficiently model humans.
We propose a deep switching state space model that can capture both state-dependent and time-dependent switching patterns in time series data.
We embed symbolic knowledge expressed as linear temporal logic (LTL) and use these embeddings to guide the training of deep sequential models.
We construct a shared latent space from different sensory modalities via contrastive learning.
Bridging the gap between symbolic and connectionist paradigms via Graph Neural Network embeddings
We present a new method for training GANs via characteristic functions
Training robots that can interactively assist humans with private information
Leveraging prior symbolic knowledge to improve the performance of deep models.
We present an approach to generate new items for groups of users based on their interaction history.
Using Bayesian Optimization to address the ill-posed nature of Inverse Reinforcement Learning
