Contact Learning, Internet-Scale Policies, and LLM Reward Engines

Contact-Rich Sim-to-Real with Minimal Real-World Data

This paper tackles the eternal sim-to-real headache for contact-rich tasks like insertion and assembly. The key insight: learning contact-aware corrections directly in real hardware with minimal real-world data shows actual industrial promise.

LLM-Driven Reward Shaping for Scalable Robot Training

A scalable way to generate robot reward functions from natural language descriptions. If you're tired of hand-tuning rewards for every new task, this is worth your lunch break.

Internet-Scale Pretraining for Zero-Shot Manipulation

Strong results on learning manipulation policies directly from internet videos without any robot-specific data. The zero-shot transfer numbers are compelling for teams who can't afford massive teleoperation setups.

Efficient Hierarchical Multi-Agent Robot Coordination

Multi-agent coordination without burning compute budgets. This hierarchical approach lets robot fleets scale without the N-squared communication explosion that kills most decentralized systems.

Robust Dexterous Manipulation with Learned Sim-to-Real Transfers

Bridges the gap between expensive sim environments and robust real-world deployment for dexterous tasks. The methodology is notably more practical than most sim-to-real papers that assume perfect domain knowledge.

Practical LLM Benchmark for Real-World Robotic Tasks

A benchmark that actually measures what matters for LLM-guided robotics beyond toy tasks. If you're evaluating foundation models for real manipulation work, this framework saves you a week of metric design.