Zero-Shot Sim, Contact Manipulation, and Safe Industrial LLMs

Research

Zero-Shot Sim-to-Real Segmentation

This looks at using a segmenter for sim-to-real which means less manual domain randomization. Useful if you need a robot to actually pick up arbitrary objects in the real world without spending weeks on data.

Typhoon: Unifying Locomotion and Communication

Getting robots to walk and talk simultaneously is hard. This framework unifies locomotion and communication, which is critical for human-robot teams where the bot needs to move through a space while coordinating tasks.

Tools

Multi-Agent Trajectory Forecasting Without the GPU Burn

If you're running multiple agents or a fleet, trajectory forecasting is usually a compute nightmare. This paper introduces multi-agent forecasting with LLMs which helps with long-horizon planning without burning through the GPU budget.

Analysis

Robust Sim-to-Real for Contact-Rich Manipulation

Handling contact-rich tasks (like inserting keys or screwing lids) in simulation often falls apart in reality. This work fixes the sim-to-real gap specifically for contact, solving a massive pain point for anyone trying to automate assembly.

Events

The Future of Space Robotics: Hardware Demos and Timelines

Space robotics usually moves slow, but this shows off real hardware demos and gives a timeline for the future. Worth watching if you need to think about autonomy in extreme environments where remote control latency kills you.

News

Safety Alliances for Industrial LLM-Robot Systems

People always forget safety when bolting LLMs onto machines. This paper lays out a safety management system specifically for Industrial LLM robotics, which is the boring-but-necessary framework you need before deploying this stuff.