Vision-Guided Manipulation, LLM Rewards, and Hierarchical Teams

Research

Vision-Guided Contact Reshaping for Deformable Manipulation

This paper introduces a method for robots to actively reshape their grip around delicate or non-rigid items using vision and touch data. This is a big step toward robots that don't just hold things, but actually manipulate them like we do.

Large Vision Models for Zero-Shot Reward Generation

The authors propose using Large Vision Models to generate dense rewards for robot learning, which massively speeds up training in simulation. This cuts out a huge manual bottleneck in getting robots to perform complex tasks.

Learning from Unlabelled Internet Videos at Scale

A practical framework for teaching robots from noisy, unstructured internet videos instead of pristine lab data. This is a scalable approach to building massive, diverse robot datasets cheaply.

Methodology

Hierarchical Correlated Equilibrium for Robot Teams

This work frames task allocation between agents as a hierarchical game, which is a robust way to handle multi-robot coordination in the wild. It keeps individual autonomy while still allowing for team-level objectives.

Analysis

Multi-Agent LLMs for Preference-Driven Control

Interesting exploration of using multi-agent LLMs as the 'executive function' for robots, letting them reason about human preferences and adapt in real-time. The key insight is separating the 'why' from the physical execution.

News

RoCoBench II: Complex Collaboration Benchmark

A new benchmark for testing how well VLMs and robots perform complex, long-horizon tasks in realistic simulations. Essential for measuring actual progress beyond simple pick-and-place demos.