Amber Xie

I am a final-year PhD student at Stanford advised by Dorsa Sadigh. I am also a research intern at Amazon FAR (Frontier AI & Robotics).

Previously, my research was supported by the NDSEG Fellowship. I was also awarded the NSF Graduate Research Fellowship. I graduated from UC Berkeley with an MS in Computer Science and BA in Computer Science and Applied Math. I was fortunate to be advised by Pieter Abbeel, Stephen James, and Youngwoon Lee at BAIR.

Contact: amberxie@stanford.edu
Google Scholar  /  Twitter  /  GitHub

I'm interested in robot learning. In particular, I would like to build intelligent, adaptable, multi-purpose robots that can learn quickly and accomplish complex tasks. I've worked on a variety of projects in deep reinforcement learning, imitation learning, and generative models to this aim.

DexPort is a portable, hand-held interface for collecting dexterous hand demonstrations. It uses the same hand hardware for collection and deployment, avoiding an embodiment gap, and records the commanded joint targets that policies need to learn forceful grasps.

Handelbot is the first learning-based system for real-world, two-handed piano playing. We train a simulation policy, and adapt it to the real world using policy refinement and residual RL.

Manipulation often requires dexterous in-hand tool use, complex contact interactions, and stability over long horizons. We propose DexDrummer, a hierarchical sim-to-real framework that integrates these capabilities within a unified drumming testbed.

Importance Weighted Retrieval (IWR) is a retrieval method that estimates importance weights, or the ratio between the target and prior data distributions, using Gaussian KDEs.

We propose Latent Diffusion Planning (LDP), a modular approach consisting of a planner which can leverage action-free demonstrations, and an inverse dynamics model which can leverage suboptimal data, that both operate over a learned latent space.

FlowRetrieval leverages optical flow representations for extracting relevant prior data and guiding policy learning to maximally benefit from the retrieved data.

We propose the domain of Language-Conditioned Path Planning (LAPP), where contact-awareness is incorporated into the path planning problem. As a first step, we propose Language-Conditioned Collision Functions (LACO).

LAMP pretrains a language-conditioned agent without human supervision using noisy VLM rewards and unsupervised reinforcement learning.

VectorFusion generates infinitely scalable vector graphics (SVGs), pixel art and sketches from text using the pretrained Stable Diffusion model.

IRM leverages the skill discriminator from unsupervised RL pretraining to perform environment-interaction-free skill sequencing for unseen downstream tasks.

Segmentations are a concise, compressed representation for images. We show sim-to-real transfer through training an RL agent on image segmentations for off-road autonomous driving.

Website template from Jon Barron