Research map / Vision-language and multimodal / Robot action and navigation
RoboWheel: A Data Engine from Real-World Human Demonstrations for Cross-Embodiment Robotic Learning
Yuhong Zhang, Zihan Gao, Shengpeng Li, Ling-Hao Chen, Kaisheng Liu, Runqing Cheng, Xiao Lin, Junjia Liu, Zhuoheng Li, Jingyi Feng, Ziyan He, Jintian Lin and 3 more
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 |
|---|---|---|---|
| add TSDF penetration optimization and RL refinement together (ablation) | RoboWheel · HO-Cap HOI reconstruction | replay success rate: 78.7 -> 93.6 | Improved |
| add RL refinement only (no TSDF) (ablation) | RoboWheel · HO-Cap HOI reconstruction | replay success rate: 78.7 -> 61.7 | Hurt |
| add TSDF only (ablation) | RoboWheel · HO-Cap HOI reconstruction | hand jitter: 3.34 -> 0.84 | Improved |
| add RL on top of TSDF (ablation) | RoboWheel · HO-Cap HOI reconstruction | object F10: 83.2 -> 89.1 | Improved |
| train on 10 HORA trajectories instead of 10 teleop demos | RDT · Real-world easy tasks | success rate (%): 66.3 -> 47.5 | Hurt |
| train on 10 HORA trajectories instead of 10 teleop demos | Pi0 · Real-world hard tasks | success rate (%): 40.0 -> 31.3 | Hurt |
| train on 10 HORA trajectories instead of teleop | Diffusion Policy · Real-world hard tasks | success rate (%): 1.3 -> 6.3 | Mixed |
| pretrain on 5k HORA trajectories before fine-tuning | Pi0 · Real-world hard tasks | success rate (%): 31.3 -> 58.8 | Improved |
| pretrain on 5k HORA trajectories before fine-tuning | RDT · Real-world hard tasks | success rate (%): 25.0 -> 47.5 | Improved |
| add simulation data augmentation (HORA-aug) (ablation) | RDT · Unseen background | success (out of 10): 1.50 -> 4.00 | Improved |
| add simulation data augmentation (HORA-aug) (ablation) | RDT · Unseen object | success (out of 10): 3.75 -> 4.25 | No clear effect |
| use RoboWheel gripper retargeting vs prior methods | RoboWheel · Robot direct replay | success rate (%): 50.0 -> 91.7 | Improved |
Closest papers by meaning
- HandEdit: A Unified Benchmark for Egocentric Human-to-Robot Dexterous Hand Image Editing (NeurIPS 2026 · similarity 0.90)
- VidBot: Learning Generalizable 3D Actions from In-the-Wild 2D Human Videos for Zero-Shot Robotic Manipulation (CVPR 2025 · similarity 0.90)
- FlowHOI: Flow-based Semantics-Grounded Generation of Hand-Object Interactions for Dexterous Robot Manipulation (NeurIPS 2026 · similarity 0.89)
- Scaffolding Dexterous Manipulation with Vision-Language Models (NeurIPS 2025 · similarity 0.89)
- Cross-Embodiment Dexterous Grasping with Reinforcement Learning (ICLR 2025 · similarity 0.89)
- HO-Cap: A Capture System and Dataset for 3D Reconstruction and Pose Tracking of Hand-Object Interaction (NeurIPS 2025 · similarity 0.89)
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.
