Research map / Vision-language and multimodal / Multimodal LLMs
3DWay: Generalizing Robot Manipulation via 3D Consistent Waypoints
Huang Ziqin, Yingyue Li, Chenyangguang Zhang, Ruida Zhang, Yuxin Chen, Gu Wang, Xingyu Liu, Masayoshi TOMIZUKA, Xiangyang Ji
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 3DWay-TD direct execution instead of pi0.5 | manipulation policy · RLBench (10 tasks, unseen) | success rate (%): 18.8 -> 64.0 | Improved |
| use 3DWay-TD instead of pi0.5 | manipulation policy · VLABench | avg success rate (%): 31.3 -> 37.7 | Improved |
| 3DWay-TD vs pi0.5 on semantic-instruction dimension | manipulation policy · VLABench semantic-instruction | success rate (%): 25.8 -> 24.3 | Hurt |
| add 3DWay waypoints (zero-shot generator) to pi0 fine-tuning | pi0 · RLBench 5 tasks, 10 demos/task | success rate (%): 17.6 -> 38.1 | Improved |
| use few-shot fine-tuned waypoint generator instead of zero-shot | pi0 + 3DWay · RLBench 5 tasks, 10 demos/task | success rate (%): 38.1 -> 46.1 | Improved |
| add 3DWay augmentation with 10 demos vs standard with 50 demos | pi0 · RLBench 5 tasks | success rate (%): 46.1 (10 demos) vs 46.7 (50 demos) | Improved |
| add 3DWay to pi0 on real-world basic tasks | pi0 · Real world basic tasks (AgileX PIPER) | avg success rate (%): 21.7 -> 65.8 | Improved |
| add multi-view consistent supervision instead of independent-view (ablation) | 3DWay generator · RLBench (unseen) | success rate (%): 22.0 -> 64.0 | Improved |
| scale backbone from NVILA-2B to 15B (ablation) | 3DWay generator · RLBench (unseen) | success rate (%): 50.1 -> 64.0 | Improved |
| perturb camera pose at test time (ablation) | 3DWay generator · RLBench (unseen) | success rate (%): 64.0 -> 61.6 (small), 58.4 (large) | Hurt |
| change robot embodiment and camera setup (ablation) | 3DWay-TD · Real world | avg success rate (%): 43.9 -> 45.0 / 45.5 / 44.7 | No clear effect |
Closest papers by meaning
- Localizing, Structuring, and Rendering: Bridging 3D and 2D Vision-Language-Action Models for Robotic Manipulation (CVPR 2026 · similarity 0.91)
- OmniManip: Towards General Robotic Manipulation via Object-Centric Interaction Primitives as Spatial Constraints (CVPR 2025 · similarity 0.89)
- MolmoMotion: Forecasting Point Trajectories in 3D with Language Instruction (NeurIPS 2026 · similarity 0.89)
- DynaRend: Learning 3D Dynamics via Masked Future Rendering for Robotic Manipulation (NeurIPS 2025 · similarity 0.88)
- GravMAD: Grounded Spatial Value Maps Guided Action Diffusion for Generalized 3D Manipulation (ICLR 2025 · similarity 0.88)
- PointZero: 3D Point Track Completion for Learning Transferable 3D Dynamics (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.
