Reproduce “KineBench: Benchmarking Embodied World Models via IDM-Free Kinematic Grounding” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zeyu Liu, Zhangzhe Zhu, Yang Zhang, Chenyou Fan, Chenjia Bai, Xuelong Li
KineBench introduces an IDM-free closed-loop framework for evaluating Embodied World Models (EWMs), mitigating attribution ambiguity by replacing learned action extraction with an explicit kinematic grounding pipeline. The framework incorporates robot-centric 3D kinematic metrics and a four-suite benchmark, revealing that current EWMs struggle with complex contact dynamics and exhibit non-uniform scaling behaviors across tasks.
Key insights
- An explicit, modular kinematic grounding pipeline significantly reduces attribution ambiguity and offers a more robust and diagnosable interface for EWM evaluation compared to learned Inverse Dynamics Models.
- Classical 3D kinematic metrics like SPARC and Manipulability Index provide specialized diagnostic signals, effectively identifying physical hallucinations, temporal fluency issues, and violations of robot kinematic constraints that pixel-based metrics miss.
- EWM performance gains from increasing data and compute are non-uniform and task-dependent, with increasing data sometimes proving detrimental for simpler tasks, challenging common scaling assumptions and highlighting the need for tailored strategies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.