Research map / Vision-language and multimodal / Robot action and navigation
When would Vision-Proprioception Policies Fail in Robotic Manipulation?
Jingxian Lu, Wenke Xia, Yuxuan Wu, Zhiwu Lu, Di Hu
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 GAP gradient adjustment to vision-proprioception policy | Concatenation policy · RoboSuite stack (MLP head) | success rate (%): 54.6 -> 77.6 | Improved |
| naively concatenate proprioception with vision | Vision-only policy · Meta-World assembly | success rate (%): 82.6 -> 74.6 | Hurt |
| add GAP to concatenation policy on out-of-distribution placement | Concatenation policy · RoboSuite threading OOD | success rate (%): 28 -> 49 | Improved |
| apply GAP during Octo fine-tuning with proprioception | Octo-VP · RoboSuite threading | success rate (%): 57 -> 78 | Improved |
| apply GAP to a UNet diffusion policy head | Concatenation policy · RoboSuite threading | success rate (%): 40 -> 63 | Improved |
| add GAP on top of Summation and FiLM fusion | Summation / FiLM fusion · Meta-World and RoboSuite (8 tasks) | success rate (%): e.g. FiLM disassemble 47 -> 85 | Improved |
| replace CPD motion-distance phases with HDBSCAN clustering (ablation) | GAP · RoboSuite threading | success rate (%): 37 (HDBSCAN) -> 52 (CPD) | Improved |
| use motion distance instead of cosine distance for CPD (ablation) | GAP · RoboSuite threading | success rate (%): 50 (CoTPC) -> 52 | Improved |
| predict transition indicator with LSTM instead of fixed rho (ablation) | GAP · put hammer into drawer / threading / cube | success rate (%): fixed rho=0.5: 89/48/75 -> LSTM: 91/52/85 | Improved |
| increase adjustment strength lambda too high (0.8) or too low (0.1) (ablation) | GAP · Meta-World assembly | success rate (%): 94 (0.3) -> 70 (0.8), 77 (0.1) | Hurt |
| apply GAP for too many epochs (90 vs 50) (ablation) | GAP · Meta-World assembly | success rate (%): 94 -> 72 | Hurt |
| add GAP to the vision branch before linear probing (ablation) | vision-proprioception policy · assembly OOD (linear probe) | success rate (%): 45 -> 69 | Improved |
Closest papers by meaning
- Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets (ICLR 2025 · similarity 0.87)
- STEP: Warm-Started Visuomotor Policies with Spatiotemporal Consistency Prediction (ICML 2026 · similarity 0.87)
- Phase-Adaptive Fusion: Spatio-Temporal Modulation for VLA Models (NeurIPS 2026 · similarity 0.87)
- Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transformers (NeurIPS 2024 · similarity 0.86)
- Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.85)
- FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency (NeurIPS 2025 · similarity 0.85)
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.
