Research map / 3D vision / 3D generation
Bootstrap Dynamic-Aware 3D Visual Representation for Scalable Robot Learning
Qiwei Liang, Boyang Cai, Minghao Lai, Sitong Zhuang, Tao Lin, Yan Qin, Yixuan Ye, Jiaming Liang, Renjing Xu
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 a diffusion transformer forward dynamics model instead of a deterministic Transformer (ablation) | AFRO · Real-world tasks | mean success rate (%): 76.25 -> 84.00 | Improved |
| add feature differencing input to the inverse dynamics model (ablation) | AFRO · Real-world tasks | mean success rate (%): 75.00 -> 84.00 | Improved |
| add inverse-consistency supervision (ablation) | AFRO · Real-world tasks | mean success rate (%): 73.75 -> 84.00 | Improved |
| use VICReg loss instead of plain MSE on features (ablation) | AFRO · Real-world tasks | mean success rate (%): 57.50 -> 84.00 | Improved |
| use latent action dimension 16 rather than 64 or 128 (ablation) | AFRO · Real-world tasks | mean success rate (%): 80.00 / 76.25 -> 84.00 | Improved |
| use frame interval k=4 rather than k=1 or k=16 (ablation) | AFRO · Real-world tasks | mean success rate (%): 81.25 / 77.50 -> 84.00 | Improved |
| pretrain on multiple task domains instead of in-domain only (ablation) | AFRO · MetaWorld and Adroit | mean success rate (%): 68.0% -> 82.8% | Improved |
| pretrain on large out-of-domain RH20T data before fine-tuning | AFRO · Real-world tasks | mean success rate (%): 70.0% -> 84% | Improved |
| pretrain on out-of-domain data for dynamics baselines such as DynaMo-3D and FVP | DynaMo-3D, FVP · Real-world tasks | mean success rate (%): DynaMo-3D 48 -> 58; FVP 34 -> 46 | Improved |
| add dynamics modeling to 3D pretraining via DynaMo-3D | static 3D pretraining · MetaWorld | mean success rate (%): PointMAE 63.9 vs DynaMo-3D 64.9; FVP 54.3 | Mixed |
| use AFRO instead of baselines for unseen object transfer | Bell Pressing · unseen objects | success rate (%): AFRO 75 -> 60 vs FVP 40 -> 5 | Improved |
| use AFRO under cluttered scenes | Bell Pressing, B2B Alignment · cluttered scenes | success rate (%): 75 -> 70; 70 -> 65 | Improved |
Closest papers by meaning
- DynaRend: Learning 3D Dynamics via Masked Future Rendering for Robotic Manipulation (NeurIPS 2025 · similarity 0.89)
- Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets (ICLR 2025 · similarity 0.89)
- VidMan: Exploiting Implicit Dynamics from Video Diffusion Model for Effective Robot Manipulation (NeurIPS 2024 · similarity 0.89)
- DynaFLIP: Rethinking Robotics Perception via Tri-Modal-Dynamics Guided Representation (NeurIPS 2026 · similarity 0.88)
- Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.88)
- PointZero: 3D Point Track Completion for Learning Transferable 3D Dynamics (NeurIPS 2026 · similarity 0.87)
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.
