TARSphere: A Robot-Oriented Full-Stack Framework for Long-Horizon Embodied Agents in Evolving Physical and Social Environments
Across eight models, general-purpose models can outperform embodied models, but changing tasks and human interaction remain difficult.
TARSphere tests robot agents in a physics simulator with embodied characters and constrained dialogue. Its benchmark spans 360 episodes across 30 indoor scenes, covering navigation-and-reporting and fetch-and-deliver tasks under five paired conditions. A frozen multimodal model uses executable skills, hierarchical memory and execution feedback to revise its decisions. Frequent collisions and falls led the authors to call for evaluating execution safety alongside task success.
HF Daily Papers' note
TARSphere tests robot agents in a physics simulator with embodied characters and constrained dialogue. Its benchmark spans 360 episodes across 30 indoor scenes, covering navigation-and-reporting and fetch-and-deliver tasks under five paired conditions. A frozen multimodal model uses executable skills, hierarchical memory and execution feedback to revise its decisions. Frequent collisions and falls led the authors to call for evaluating execution safety alongside task success.
HF Daily Papers' note
score 5