Research map / Graph learning / Graph self-supervised learning
SIR: Structured Image Representations for Explainable Robot Learning
Paul Mattes, Jan Schwab, Jens Bosch, Maximilian Xiling Li, Nils Blank, Minh-Trung Tang, Moritz Haberland, Rudolf Lioutikov
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 fully-connected scene graph instead of image embedding | MDT · RoboCasa 24 tasks | success rate: 14.81 -> 16.98 | Improved |
| add learned instruction-grounded sparsification to FC graph (ablation) | MDT graph policy · RoboCasa 24 tasks | success rate: 16.98 -> 19.50 | Improved |
| graph representation for Drawers and Pick/Place tasks | MDT · RoboCasa Drawers | success rate: 49.75 -> 46.25 | Hurt |
| graph representation for Levers | MDT · RoboCasa Levers | success rate: 23.67 -> 48.50 | Improved |
| add soft histogram loss to node removal (ablation) | SIR sparsification · RoboCasa 24 tasks | success rate: 9.60 -> 19.50 | Improved |
| random node removal instead of learned (ablation) | SIR sparsification · RoboCasa 24 tasks | success rate: 19.50 -> 5.48 | Hurt |
| use task-specific k and L1 node-weight guidance (SIR vs TopK) (ablation) | SIR sparsification · RoboCasa 24 tasks | success rate: 18.44 -> 19.50 | Improved |
| use threshold sparsification instead of FC (ablation) | FC graph · RoboCasa 24 tasks | success rate: 16.98 -> 17.17 | No clear effect |
| cropped image features as node features vs label+bbox (ablation) | Split-View graph · RoboCasa 24 tasks | success rate: 10.98 -> 16.65 | Improved |
| use point clouds via GNN nodes instead of directly into action generator (ablation) | MDT · RoboCasa 24 tasks | success rate: 4.13 -> 11.08 | Improved |
| introduce 3-9 distractor objects at inference (ablation) | Image baseline · RoboCasa with distractors | success rate: drop of 3.3% on average | Hurt |
| introduce distractor objects at inference (ablation) | SIR / FC-Graph · RoboCasa with distractors | success rate: almost no degradation | No clear effect |
Closest papers by meaning
- GraphMimic: Graph-to-Graphs Generative Modeling from Videos for Policy Learning (CVPR 2025 · similarity 0.88)
- Action Images: End-to-End Policy Learning via Multiview Video Generation (NeurIPS 2026 · similarity 0.87)
- DiffusionVLA: Scaling Robot Foundation Models via Unified Diffusion and Autoregression (ICML 2025 · similarity 0.87)
- Prediction with Action: Visual Policy Learning via Joint Denoising Process (NeurIPS 2024 · similarity 0.87)
- Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets (ICLR 2025 · similarity 0.86)
- DMC-VB: A Benchmark for Representation Learning for Control with Visual Distractors (NeurIPS 2024 · 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.
