Research map / Image generation / Autoregressive and latent generation
DynBridge: Bridging Imagination and Control through Interaction Dynamics for Robot Manipulation
Alex Wang, Zhiwei Dong, Qicheng Bai, Chenshi Zhang, Yujie Yi, Guang Dai, Yong Liu, Mengmeng Wang
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 |
|---|---|---|---|
| train the interaction dynamics generator and action predictor jointly end-to-end instead of separately (ablation) | DynBridge · LIBERO-Spatial, Goal, Long | success rate | Improved |
| add trajectory reconstruction supervision alongside action imitation (ablation) | DynBridge · LIBERO-Spatial, Goal, Long | success rate | Improved |
| use latent interaction representation instead of explicit trajectory coordinates (ablation) | DynBridge · LIBERO-Spatial, Goal, Long | success rate | Improved |
| predict future trajectories (horizon L greater than 0) instead of none (ablation) | DynBridge · LIBERO suites | success rate | Improved |
| increase trajectory prediction horizon to a very long length (ablation) | DynBridge · LIBERO suites, including LIBERO-Long | success rate | Hurt |
| add the action-conditioned dynamics aggregator across dynamic token counts 4, 8, 16, 32 (ablation) | DynBridge · LIBERO | success rate | Improved |
| increase number of dynamic tokens without aggregator (ablation) | DynBridge w/o aggregator · LIBERO | success rate | Mixed |
| condition aggregator on action tokens rather than vision, language or learnable anchor (ablation) | DynBridge aggregator · LIBERO-Spatial, Goal, Long | success rate | Improved |
| add any aggregator conditioning (vision, language, learnable, action) versus none (ablation) | DynBridge · LIBERO-Spatial, Goal, Long | success rate | Improved |
| add interaction dynamics modeling for unseen goal positions (ablation) | DynBridge · modified LIBERO-Goal, displaced targets | success rate | Improved |
| add interaction dynamics modeling in real-world tasks (ablation) | DynBridge · Franka Research 3, five real tasks, 10 trials | success rate: +22% | Improved |
| transfer Panda-learned dynamics to XArm7 with imperfect demonstrations | DynBridge · XArm7 cross-robot transfer | success rate | Improved |
Closest papers by meaning
- DynScene: Scalable Generation of Dynamic Robotic Manipulation Scenes for Embodied AI (CVPR 2025 · similarity 0.89)
- DreamVLA: A Vision-Language-Action Model Dreamed with Comprehensive World Knowledge (NeurIPS 2025 · similarity 0.88)
- Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.88)
- Chain-of-Action: Trajectory Autoregressive Modeling for Robotic Manipulation (NeurIPS 2025 · similarity 0.88)
- DynaFLIP: Rethinking Robotics Perception via Tri-Modal-Dynamics Guided Representation (NeurIPS 2026 · similarity 0.88)
- SyncWorld: Visual Calibration Enables World Models as Zero-Shot Simulators (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.
