Jonathan Chang, David Goedicke, Natalie Friedman, Travers Rhodes, PhD students, Cornell Tech
Location: 122 Gates Hall
Jonathan Chang: Mitigating Covariate Shift in Imitation Learning
Covariate shift is a core issue in Imitation Learning (IL). Traditional IL methods like behavior cloning (BC) (Pomerlau, 1989), while simple, suffer from covariate shift, learning a policy that can make arbitrary mistakes in parts of the state space not covered by the expert dataset. This leads to compounding errors in the agent’s performance (Ross and Bagnell, 2010), hurting the generalization capabilities in practice.
In this talk, I will present our recent work studying offline Imitation Learning (IL) where an agent learns to imitate an expert demonstrator without additional online environment interactions. Instead, the learner is presented with a static offline dataset of state-action-next state transition triples from a potentially less proficient behavior policy. We introduce Model-based IL from Offline data (MILO): an algorithmic framework that utilizes the static dataset to solve the offline IL problem efficiently and mitigate this covariate shift phenomenon.
Natalie Friedman: The Functions of Clothes For Robots
Most robots are unclothed. I believe that robot clothes present an underutilized opportunity for the field of designing interactive systems. Clothes can help robots become better robots––by helping them be useful in a new, wider array of contexts, or better adapt and function in the contexts they are already in. To make clothes for robots, I am learning how to drape fabric onto robots from Kari Love, a Broadway costumer. In this lightning talk I will share our process, including swatching and draping on a Kinova Gen 3 and Blossom.
David Goedicke: Imagining Future Automations with VR
I build specialized Virtual Reality Simulators that allow us to assess specific interactions between people and machines. Many of these focus on Autonomous Vehicles; recent projects started to focus on integrating ROS2 into these simulations to test and validate programmed robotic behaviors in VR before deploying them on any robot.
Travers Rhodes: Local Disentanglement in Variational Auto-Encoders Using Jacobian L1 Regularization
There have been many recent advances in representation learning; however, unsupervised representation learning can still struggle with model identification issues. Variational Auto-Encoders (VAEs) and their extensions such as Beta-VAEs have been shown to locally align latent variables with PCA directions, which can help to improve model disentanglement under some conditions. We propose adding an L1 loss (sparsity cost) to the VAE’s generative Jacobian during training to encourage local latent variable alignment with independent factors of variation in the data. I’ll present qualitative and quantitative results that show our added L1 cost encourages local axis alignment of the latent representation with individual factors of variation.