Multi-Modal Learning for Multi-Suction Picking at Scale
We learned to score multi-suction robot picks from real industrial data using visual pretraining and cross-modal attention.
AI · Robotics · Embodied Intelligence
Dylan Randle.
Applied scientist and technical lead building learning systems for dexterous robot manipulation.
I work across imitation learning, reinforcement learning, perception, and deployed robotics—turning research into systems that operate at industrial scale.
Projects & Research
Learning Object Properties Through Robot Proprioception
We used differentiable robot-object simulations to estimate mass and softness from joint encoders alone.
MuST: A Multi-Head Skill Transformer for Long-Horizon Manipulation
We developed a transformer policy that learns reusable manipulation skills and chains them using estimates of skill progress.
Rubik's Cube Solving Robot
A robot that autonomously solves a Rubik's cube.
Avoiding Object Damage in Robotic Manipulation
We built and deployed a damage-risk classifier that reduced robot-induced object damage in a warehouse manipulation system.
Real-Time Quantitative Trading System
I built a live trading system that used topic models over financial news to predict cross-sectional price movements at the market open.
Golf Swing Analysis with Computer Vision
I explored how well human pose estimation could track a golfer's body through a fast, highly articulated swing.
Generating Faces with a ResNet VAE
I trained a variational autoencoder (VAE) on the CelebA dataset to generate faces.
Physics-Informed Generative Adversarial Networks for Differential Equations
We developed an adversarial training objective that lets physics-informed neural networks learn how to penalize differential-equation residuals.
Differentiable Neural Architecture Search for Scientific Datasets
We evaluated DARTS as an efficient way to discover neural architectures for materials science, astronomy, and medical imaging.
Interpretable Reinforcement Learning for Healthcare
We distilled black-box reinforcement-learning policies into readable decision sets and measured the tradeoff between interpretability and performance.
Causal LSTMs for Mouse Microbiome Modeling
We used sparse LSTM models and Granger-causal analysis to compare microbial interactions in healthy and IBD mice.
Distributed YouTube-8M Training with Spark and TensorFlow
We trained bidirectional LSTM video classifiers over YouTube-8M using Spark, Elephas, HDFS, and an AWS EMR cluster.
Automatic Differentiation from Scratch
We built a Python library for forward- and reverse-mode automatic differentiation, Jacobians, and gradient-based optimization.
Detecting Twitter Trolls with Machine Learning
We compared text representations and classifiers for identifying tweets from Internet Research Agency accounts.
Bayesian Generative Adversarial Networks
We reproduced and evaluated a Bayesian GAN that uses stochastic-gradient MCMC to improve mode diversity and semi-supervised learning.