We propose Curriculum-RLAIF, a data-centric framework that improves reward model generalizability by training on preference pairs of increasing difficulty. This curriculum-based approach addresses data noise, distribution shift, and model-capacity mismatch. Experiments show that Curriculum-RLAIF significantly boosts policy alignment performance without extra inference cost, outperforming non-curriculum and alternative strategies.
This study explores whether LLMs can mentally model decision-making agents by reasoning over their behavior and state transitions from interaction histories. Evaluated on reinforcement learning tasks, results show that while LLMs offer some insight, they fall short of fully modeling agents without further innovation, highlighting both their potential and current limitations for explainable RL.
We propose REAL (Response Embedding-based Alignment for LLMs), a method to improve alignment efficiency by selecting less ambiguous, dissimilar response pairs for annotation. By leveraging embedding similarity in an off-policy manner, REAL reduces label noise and improves alignment quality. Experiments show it boosts performance while cutting annotation effort by up to 65%.
LLM+MAP is a bimanual planning framework that combines GPT-4o with multi-agent task planning to enable efficient and logically consistent long-horizon manipulation. It outperforms baseline LLMs on planning time, success rate, and coordination metrics.
We propose an LLM-driven framework that enables **robots to autonomously discover useful skills from scratch**. By generating tasks, rewards, and success criteria, the LLM guides reinforcement learning, while a vision-language model verifies outcomes. This allows the robot to build a meaningful skill library without relying on predefined primitives.
We propose LoT (Logical Thoughts), a framework that improves large language models’ reasoning at inference time by applying symbolic logic to verify and correct their step-by-step thought process. LoT enhances performance on diverse reasoning tasks and reduces hallucinations.
We introduce OSSA (Object State-Sensitive Agent), a task-planning agent using pre-trained LLMs and VLMs to generate plans sensitive to object states. We compare two methods: a modular approach combining vision and language models, and a monolithic VLM approach. Evaluated on tabletop tasks involving clearing a table, OSSA’s monolithic model outperforms the modular one. A new multimodal benchmark dataset with object state annotations is provided.
LABOR uses LLMs to orchestrate control policies for long-horizon bimanual manipulation tasks. By leveraging task reasoning and coordination via language, it achieves higher success rates on simulated tasks with the NICOL robot and provides insights into LLM-based control challenges.
We present Matcha agent, an interactive perception framework that uses LLMs to guide robots in gathering multimodal sensory data (vision, sound, haptics, proprioception) before executing tasks. Matcha enables high-level reasoning and planning in partially observable environments, showing that LLMs can effectively control robot behavior when grounded with multimodal context.