Research map / Video / Video generation
KineBench: Benchmarking Embodied World Models via IDM-Free Kinematic Grounding
Zeyu Liu, Zhangzhe Zhu, Yang Zhang, Chenyou Fan, Chenjia Bai, Xuelong Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LoRA instead of full-parameter fine-tuning | Wan2.2 5B · KineBench Suite 0 | success rate (%): 56.32 -> 19.00 | Hurt |
| use LoRA instead of full-parameter fine-tuning | Wan2.2 5B · KineBench Suite 1 | success rate (%): 11.90 -> 16.67 | Improved |
| use LoRA instead of full-parameter fine-tuning | Wan2.2 5B · KineBench Suite 2 (unseen assets) | success rate (%): 55.50 -> 26.33 | Hurt |
| increase training steps from 1.5k to 15k (ablation) | Wan2.1 1.3B · KineBench Suite 3 | success rate (%): 44.83 -> 73.33 | Improved |
| increase expert trajectories from scale 10 to scale 50 (ablation) | Wan2.1 1.3B · KineBench Suite 3 | success rate (%): 53.67 -> 51.17 | No clear effect |
| increase homogeneous training trajectories from 10 to 100 (ablation) | Wan2.1 1.3B · KineBench Suite 3, OpenBoxEasy | success rate (%): 100% to 83.33% | Hurt |
| increase optimization budget (ablation) | Wan2.1 1.3B · KineBench Suite 3, StoreCube-v2 | success rate (%): 13.34% to 69.52% | Improved |
| evaluate on unseen assets instead of seen assets | Wan2.2 · KineBench Suite 2, OpenBoxHard | success rate (%): 60.0% to 30.0% | Hurt |
| replace ground-truth depth with MoGeV2-estimated depth (ablation) | FoundationPose pipeline · KineBench pipeline pose accuracy | translational error: roughly 1.5--3 cm | Hurt |
| use explicit KineBench pipeline instead of IDM | action extraction · Unseen simulator trajectories | position error: order of ten centimeters | Improved |
Closest papers by meaning
- WorldSimBench: Towards Video Generation Models as World Simulators (ICML 2025 · similarity 0.90)
- EnerVerse: Envisioning Embodied Future Space for Robotics Manipulation (NeurIPS 2025 · similarity 0.89)
- RoboScape: Physics-informed Embodied World Model (NeurIPS 2025 · similarity 0.89)
- Universal Actions for Enhanced Embodied Foundation Models (CVPR 2025 · similarity 0.88)
- LongScape: Advancing Long-Horizon Embodied World Models with Context-Aware MoE (NeurIPS 2026 · similarity 0.88)
- SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
