Alison Bartsch

I am a Senior Computer Vision and Machine Learning Applied Scientist at Dandy.

I recieved my PhD in Mechanical Engineering at Carnegie Mellon University's Mechanical and AI Lab (MAIL) with a focus on using machine learning for complex and dexterous robotic tasks, such as deformable object manipulation.

I recieved my BS in Aeronautics and Astronautics at Stanford University, and was an undergraduate researcher in the Autonomous Systems Lab (ASL).

Outside of my research, I have a passion for baking, collecting/propagating rare tropical plants, and playing/watching soccer.

Email  |  CV  |  Scholar  |  Github

Research

A robotic pinch pottery framework with task progress-guided 3D diffusion policy with pre-trained point cloud embeddings.

A text-to-3D real-world robotic coarse-to-fine framework for creating shapes without ground truth 3D goals.

Leveraging the world knowledge of Large Language Models to generate coherent action sequences for 3D deformable object shaping.

Goal-conditioned 3D diffusion policy with PointBERT point cloud embeddings for imitation learning for the autonomous clay sculpting task.

PointBERT model pre-trained on the ShapeNet dataset provides a quality latent embedding of 3D geometrical features of clay relevant to predicting the deformable dynamics given a grasp action.

Incorporating affordance predictions into an LLM planner pipeline for long-horizon tool use tasks.

Leveraging large pretrained Vision-Language-Action models' embedding space for fast lookup-based imitation.

Interactive perception framework for vision-based fluid viscosity prediction.

Vision-based autonomous pouring pipeline for transparent container detection, liquid segmentation, and foam formation prediction.

Leveraging foundational vision models SAM and YOLO for identifying fruits and vegetables in all states of chopped (whole to diced) for a fully autonomous chopping pipeline.

An augmentation strategy that maintains the advantages of seeding RL with human demonstrations, while only requiring a single human demonstration.

An open-source oculus-based teleoperation system for the Franka manipulator.

Factorizing large N-dimensional action spaces into N 1-dimensional action spaces speeds up and improves policy learning for Deep Q-Learning.

Education

Awards

This template is taken from Jon Barron's website's source code