Explore cutting-edge research in embodied intelligence through this seminar featuring presentations on diverse topics. Dive into automated optimal experiment design, neuro-symbolic generative models for combinatorial symbols, and meta-learning of conserved quantities. Discover set-based state estimation techniques, end-to-end differentiable frameworks for contact-aware robot design, and audio spectrogram transformers. Learn about few-shot program induction and synthesis, context features in transformer language models, and temporal and object quantification networks. Examine self-supervised speech recognition and synthesis, as well as physically-grounded abstract social events for machine social perception. Gain insights into real-world applications, including PID heater controller optimization, morphology parameterization for articulated robots, and cross-lingual mask transfer in low-resource ASR.
EI Seminar - Recent Papers in Embodied Intelligence
Massachusetts Institute of Technology via YouTube
Overview
Syllabus
Intro
Problem Setup
Real-World PID Heater Controller Optimization
Measurement = measurement consistent set (blue outline)
What is a Good Morphology Parameterization?
Deformation-based Hierarchical Morphology Parameterization Parameterize Articulated Robots
Real Experiments Test manufacturability and functionality of optimized designs
Tasks come from real codes in programming competitions Programming competitions: hundreds of correct implementations for each od -3k problems
Transformer baseline
Language modeling with transformer
Feature Representation
Unstructured Magnitude Pruning
Low-Resource (English) ASR
Cross-Lingual Mask Transfer
PHASE combines physical scene understanding and social scene understanding
Taught by
MIT Embodied Intelligence
