Research map / Vision-language and multimodal / Robot action and navigation
Disentangled Robot Learning via Separate Forward and Inverse Dynamics Pretraining
Wenyao Zhang, Bozhou Zhang, Zekun Qi, Wenjun Zeng, Xin Jin, Li Zhang
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 pretraining to forward dynamics model (ablation) | DeFI · CALVIN ABC-D (multi-view) | avg. length: 3.28 -> 4.51 | Improved |
| add pretraining to inverse dynamics model (ablation) | DeFI · CALVIN ABC-D (multi-view) | avg. length: 4.16 -> 4.51 | Improved |
| add human videos to pretraining (ablation) | DeFI · CALVIN ABC-D (multi-view) | avg. length: 3.92 -> 4.51 | Improved |
| add human videos to forward model only (ablation) | DeFI · CALVIN ABC-D (multi-view) | avg. length: 4.19 -> 4.51 | Improved |
| add human videos to inverse model only (ablation) | DeFI · CALVIN ABC-D (multi-view) | avg. length: 4.34 -> 4.51 | Improved |
| use five denoising steps instead of one (ablation) | DeFI · CALVIN ABC-D (multi-view) | avg. length: 4.51 -> 4.45 | No clear effect |
| replace SVD video backbone with DINO generative model (ablation) | DeFI · CALVIN ABC-D (multi-view) | avg. length: 4.51 -> 3.70 | Hurt |
| use VQ causal-transformer inverse model instead of MLP (ablation) | DeFI · CALVIN ABC-D (multi-view) | avg. length: 3.42 -> 4.51 | Improved |
| use VQ causal-transformer inverse model instead of plain transformer (ablation) | DeFI · CALVIN ABC-D (multi-view) | avg. length: 4.22 -> 4.51 | Improved |
| keep GFDM and GIDM trained with adapter instead of adapter only (ablation) | DeFI · CALVIN ABC-D (multi-view) | avg. length: 4.33 -> 4.51 | Improved |
| add GIDM finetuning to GFDM+adapter (ablation) | DeFI · CALVIN ABC-D (multi-view) | avg. length: 4.35 -> 4.51 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 4 of 4 on most shared results.
- TraceVLA — TraceVLA: Visual Trace Prompting Enhances Spatial-Temporal Awareness for Generalist Robotic Policies (ICLR 2025) · better on 2 of 2 shared results
- UP-VLA — UP-VLA: A Unified Understanding and Prediction Model for Embodied Agent (ICML 2025) · better on 1 of 1 shared result
- Seer — SEER: Backdoor Detection for Vision-Language Models through Searching Target Text and Image Trigger Jointly (AAAI 2024) · better on 1 of 1 shared result
- CLOVER — Clover: Towards a Unified Video-Language Alignment and Fusion Model (CVPR 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DiT4DiT: Jointly Modeling Video Dynamics and Actions for Generalizable Robot Control (NeurIPS 2026 · similarity 0.92)
- Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.91)
- Learning an Actionable Discrete Diffusion Policy via Large-Scale Actionless Video Pre-Training (NeurIPS 2024 · similarity 0.90)
- Latent Action Pretraining from Videos (ICLR 2025 · similarity 0.90)
- PlaySlot: Learning Inverse Latent Dynamics for Controllable Object-Centric Video Prediction and Planning (ICML 2025 · similarity 0.89)
- Video Prediction Policy: A Generalist Robot Policy with Predictive Visual Representations (ICML 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.
