Many-Shot Imitation, Robust Contact, and Eidetic Fleet Coordination

Few-Shot Imitation Learning for Mobile Manipulation

Demonstrates a new method to rapidly teach robots complex tasks using minimal human input. This is huge for scaling robot teams because it drastically reduces the bottleneck of data collection and supervision.

Generalizable Sim-to-Real for Contact-Rich Manipulation

Tackles the tricky problem of making policies trained in simulation work reliably in the real world, especially for force-sensitive tasks. It's a key step toward deploying lab-bots in messy warehouses and factories.

Multi-Robot Task Planning with LLM-based Agents

Introduces a framework for multiple robots to reason and plan together using large language models. This is the kind of high-level coordination needed for building true autonomous fleets, not just single-task bots.

Learning from Demos with Multimodal Sensing

Presents a system that lets robots learn precise visual and tactile skills from watching humans perform tasks loosely. The potential is turning any recorded human action into executable, robust robot programming.

Rethinking Evaluation in Robot Learning Benchmarks

Proposes a new way to evaluate robot learning benchmarks to focus on real-world applicability. Most benchmark scores are vanity metrics; this paper tries to make them actually mean something for deployment.

Safe Navigation in Human-Crowded Environments

Explores how robots can safely operate in dynamic environments filled with unpredictable human motion. Safety and robustness in human-shared spaces are non-negotiables for any commercial deployment.