Our Projects
Embodied First Exam
Embodied First Exam is the first exam for code agents in the physical world. In every task, a coding agent writes a program, runs it on a robot in a frozen scene, and sees what happens. Tasks come from many benchmark suites, including nine suites written for this exam, from robot painting and dexterous hand manipulation to robot sports matches and embodied safety.
ViewAgent
ViewAgent studies whether VLM agents can plan camera moves to find a target view in real 3D scenes. Its ViewSuite benchmark exposes a multi-turn view-planning gap and shows how on-policy view graph distillation improves spatial planning.
BAGEN
BAGEN (Budget-Aware Agent) studies whether LLM agents know what they will spend: while working on a task, a budget-aware agent predicts bounds on its remaining tokens, time, and money and stops early when the task is unlikely to finish within budget. Frontier agents turn out to be over-optimistic, and budget-aware early stopping saves 28-64% of tokens at a 1.6-4.2% success cost. Presented at the ICML 2026 RLxF and Agents in the Wild workshops.
RAGEN
We introduce RAGEN to train LLM reasoning agents via RL in multi-turn, stochastic environments. RAGEN is formulated with MDP and optimized through Reasoning-Interaction Chain Optimization (RICO). RAGEN-0.5B is trained across three agentic tasks, showing intriguing reasoning patterns.
Embodied Agent Interface
Current evaluations of LLMs in embodied AI lack standardization and detailed error analysis. Our introduce a unified interface (Embodied Agent Interface) for diverse tasks and LLM modules (planning, decomposition, etc.) and fine-grained metrics (identifying hallucination, affordance errors, etc.). This enables systematic assessment, pinpointing specific LLM limitations and strengths to inform more effective integration into embodied agents.
