Meta's Physical AI Push, Budget Teleop, and Real-Time LLM Planning

Research

Watch and Learn: Demo-free Manipulation via Human Videos

Learning manipulation from human video without needing robot-specific demonstrations is huge. It cuts the cost of training data and suggests we're getting closer to general-purpose imitation.

Fast LLM-based Robot Task and Motion Planning

Using LLMs to plan robot motion is powerful but slow. This paper makes it fast enough for real-time use in dynamic settings.

Decoupled 3D Match: Vision-Based Quantified Manual Control

Teleop usually needs expensive sensors. This is a budget way to collect training data using just a camera.

Tools

MiRA: Fast Long-Context Robotics Memory

This is a fast memory-based context system, which solves the "lost-in-the-middle" issue for multi-step robots. If you're building a bot with long-horizon planning, take a look at this.

Latent Motion Diffusions for Legged Mobility

This framework helps build more general and efficient robot walking and moving behaviors. It's a useful jump for quadruped or legged development.