TARSphere: A Robot-Oriented Full-Stack Framework for Long-Horizon Embodied Agents in Evolving Physical and Social Environments
- Published
- Source
- arXiv
- Paper number
- 1361
- Field
- Robotics
- arXiv ID
- 2610.30043
Key points
- TARSphere-Infra is the execution infrastructure that runs robot policies in a physics simulator together with embodied NPCs and constrained dialogue.
- TARSphere-Harness is the embodied brain harness that lets a frozen MLLM revise decisions using multi-level executable skills, hierarchical memory, and execution feedback.
- Agent-in-the-loop curation built TARSphere-Bench with 360 episodes across 30 indoor scenes, and five paired conditions assessed perceptual robustness, disruption adaptation, task update, human coordination, and intent understanding.
- Across 8 MLLMs, general-purpose models can outperform embodied models, and Gemini-3.8-Flash achieved 40.6% navigation and 47.8% delivery success.
- Across model-task groups, collisions occur in 84-100% of episodes and falls in 9-47%, motivating evaluation of execution safety alongside task success.
- The full curation pipeline yielded 96.7% usable episodes, and removing multi-agent review dropped this to 45.0%, the largest decline.
Paper links
External research summaries. These are not HDATF publications or measured product results.