Event

Proposal of Long-Horizon Embodied Agent Framework TARSphere and Benchmark

1 reports · 1 sources

Yicheng Xu et al. propose TARSphere, a full-stack robotics framework for long-horizon embodied agents, and construct the evaluation benchmark TARSphere-Bench, evaluating 8 multimodal large language models (MLLMs). Results show that general-purpose models can outperform embodied specialized models, but task updates and human-robot interaction remain difficult, and collisions and falls occur frequently. The framework consists of three parts: TARSphere-Infra runs robot policies in physical simulation, including embodied NPCs and constrained dialogue; TARSphere-Harness enables frozen multimodal large language models to make decisions through multi-layer executable skills; TARSphere-Bench covers 30 indoor scenes and 360 episodes. The authors state that the evaluation results also indicate that even with a shared action interface and decision budget, execution safety should be evaluated beyond task success rate.