Research map / Video / Video segmentation and tracking
Spatiotemporal Predictive Pre-training for Robotic Motor Control
Jiange Yang, Bei Liu, Jianlong Fu, Bocheng Pan, Gangshan Wu, Limin 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 |
|---|---|---|---|
| add temporal prediction with future frame (STP vs MAE) | MAE pre-training · Single-task simulation | weighted average: 59.6 -> 63.7 | Improved |
| add ImageNet-MAE hybrid initialization | STP · Single-task simulation | weighted average: 63.7 -> 64.2 | Improved |
| add temporal prediction (STP vs MAE) | MAE pre-training · LIBERO-LONG | success rate: 37.3 -> 41.3 | Improved |
| add temporal prediction (STP vs MAE) | MAE pre-training · RLbench | success rate: 45.2 -> 51.3 | Improved |
| use STP post-pre-training instead of MAE post-pre-training | post-pre-training · Single-task simulation | weighted average: 72.5 -> 76.4 | Improved |
| end-to-end fine-tuning instead of frozen encoder | STP · Single-task simulation | weighted average: 63.7 -> 62.9 | Hurt |
| scale ViT-B to ViT-L with post-pre-training | STP · Single-task simulation | weighted average: 76.4 -> 78.4 | Improved |
| STP vs VC-1 in real world standard environment | STP · Real-world | success rate: 42.0 -> 49.3 | Improved |
| STP vs VC-1 in unseen environment with distractors | STP · Real-world | success rate: 36.7 -> 46.3 | Improved |
| add spatial prediction on the 75% masked current frame (ablation) | STP · Single-task simulation | weighted average: 57.4 -> 63.7 | Improved |
| use future frame 95% masked condition instead of language only (ablation) | STP · Single-task simulation | weighted average: 55.4 -> 63.7 | Improved |
| add language narration to 95% masked future frame condition (ablation) | STP · Single-task simulation | weighted average: 63.7 -> 63.1 | No clear effect |
Closest papers by meaning
- Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets (ICLR 2025 · similarity 0.89)
- DynaMo: In-Domain Dynamics Pretraining for Visuo-Motor Control (NeurIPS 2024 · similarity 0.89)
- Video Prediction Policy: A Generalist Robot Policy with Predictive Visual Representations (ICML 2025 · similarity 0.88)
- Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.88)
- Visual Representation Learning with Stochastic Frame Prediction (ICML 2024 · similarity 0.87)
- Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers (NeurIPS 2024 · 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.
