Research map / Reinforcement learning / Visual and representation RL
Rethinking Camera Choice: An Empirical Study on Fisheye Camera Properties in Robotic Manipulation
Han Xue, Nan Min, Xiaotong Liu, Wendi Chen, Yuan Fang, Jun Lv, Cewu Lu, Chuan Wen
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 wrist fisheye (235 deg) instead of pinhole (90 deg), single camera, feature-poor | Diffusion Policy · Sim 6-task average | success rate: 0.31 -> 0.57 | Improved |
| use fisheye instead of pinhole, double camera, feature-poor | Diffusion Policy · Sim 6-task average | success rate: 0.38 -> 0.72 | Improved |
| feature-rich instead of feature-poor background (ablation) | Fisheye single camera · Sim 6-task average | success rate: 0.57 -> 0.66 | Improved |
| feature-rich instead of feature-poor background (ablation) | Fisheye double camera · Sim 6-task average | success rate: 0.72 -> 0.75 | No clear effect |
| feature-rich instead of feature-poor background (ablation) | Pinhole single camera · Sim 6-task average | success rate: 0.31 -> 0.34 | No clear effect |
| feature-rich instead of feature-poor background (ablation) | Fisheye · Real-world mean normalized score | normalized score: +0.39 vs +0.18 for pinhole | Improved |
| fisheye instead of pinhole in feature-rich scene (ablation) | CLIP encoder probe · Pick Cup proprioception | translation error (cm): 5.367 -> 2.362 | Improved |
| feature-rich instead of poor background for pinhole (ablation) | CLIP encoder probe · Fold Towel proprioception | translation error (cm): 4.204 -> 5.329 | Hurt |
| increase number of training scenes (1 to 8) (ablation) | Fisheye policy · Real Pick Cup unseen scenes | normalized score: exceeds 95% with eight scenes | Improved |
| increase number of training scenes (1 to 32) (ablation) | Fisheye vs pinhole policy · Sim Coffee unseen scenes | success rate: no numbers given | Improved |
| add Random Scale Augmentation (U(0.7,1.3)) in place of fixed-scale crop (ablation) | Diffusion Policy fisheye · Sim unseen camera params, 6-task avg | success rate | Improved |
Closest papers by meaning
- Data Scaling Laws in Imitation Learning for Robotic Manipulation (ICLR 2025 · similarity 0.88)
- Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets (ICLR 2025 · similarity 0.87)
- SAM2Act: Integrating Visual Foundation Model with A Memory Architecture for Robotic Manipulation (ICML 2025 · similarity 0.86)
- Predictive Inverse Dynamics Models are Scalable Learners for Robotic Manipulation (ICLR 2025 · similarity 0.86)
- What's the Move? Hybrid Imitation Learning via Salient Points (ICLR 2025 · similarity 0.86)
- DynaFLIP: Rethinking Robotics Perception via Tri-Modal-Dynamics Guided Representation (NeurIPS 2026 · similarity 0.86)
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.
