Zhejiang's TARSphere Benchmarks Long-Horizon Embodied Agents Across 360 Episodes, 30 Scenes
Yicheng Xu · hf · 2026-10-24
Zhejiang University and Fudan researchers released TARSphere (arXiv 2610.30043, open source), a robot-oriented full-stack framework for long-horizon embodied agents. TARSphere-Infra runs robot policies in a physics simulator with embodied NPCs and constrained dialogue, while the Harness lets a frozen MLLM act via multi-level executable skills with hierarchical memory and execution feedback. TARSphere-Bench spans 360 episodes across 30 indoor scenes under five paired conditions. Evaluating 8 MLLMs shows general-purpose models can beat embodied-specific ones, while task updates and human interaction remain hard; frequent collisions motivate evaluating execution safety alongside task success.
More from Embodied
- Rodney Brooks on Amazon's 1M robots: real-world deployment is the hard part — rodneyabrooks · 2026-10-24
- Freeform Preference Learning Lets Robot Annotators Compare Trajectories Along Natural-Language Axes — micoolcho · 2026-10-24
- The Bitter Lesson Hits Robotics: LLM Scores Double Every 54 Days as Speed Rises 4.4x — Miles_Brundage · 2026-10-24
- Builder pairs Grok Bot for firmware with locally-run Hermes Agent inside a Stack-chan robot — alexcovo_eth · 2026-10-24
- Chinese-built badminton robot tracks shuttlecock and returns shots in demo — LexiLove · 2026-10-24
- Nvidia CEO shows off 'Microducks' mini robot by Hugging Face and Pollen, under $400 — LexiLove · 2026-10-24