Data Flywheels, Safe LLMs, and Industrial Humanoids

Research

Proactive Robot Assistance: Learning to Ask for Help to Bootstrap Data

This paper tackles the core bottleneck in real-world robot learning: limited data. By teaching robots to proactively ask for help, it creates a natural data flywheel. A smart path to bootstrapping capable systems without massive initial datasets.

PRISM: A Framework for Safe LLM-Driven Robotic Control

A practical framework for safely integrating LLMs into robotic control loops. It uses formal methods to provide safety guarantees, which is the missing piece for deploying powerful language-driven agents in the physical world. Crucial for industrial adoption.

Learning Collaborative Manipulation from Human-Robot Interaction Data

Focuses on the messy real-world problem of robots learning to manipulate objects by, and with, humans. The emphasis on using novel data collection paradigms is key—it acknowledges that pure simulation data still hits walls for dexterous, collaborative tasks.

Tools

DiffPDE: Using Diffusion Models to Solve Physical Equations for Robot Design

Applying diffusion models to solve PDEs for physical system design. This isn't just a robotics paper; it's a potential game-changer for rapid prototyping of robot mechanics, materials, and control systems, cutting down simulation time radically.