Research map / Continual learning / Domain and prompt-based
GeCo-SRT: Geometry-aware Continual Adaptation for Cross-Task Sim-to-Real Transfer
Wenbo Yu, Wenke Xia, Weitao Zhang, 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 |
|---|---|---|---|
| replace Transic backbone with Geo-MoE under standard PER (ablation) | continual sim-to-real · 4-task continual | avg success rate: 40.0% -> 55.7% | Improved |
| replace standard PER with Geo-PER (expert-utilization prioritization) (ablation) | Geo-MoE · 4-task continual | avg success rate: 55.7% -> 63.3% | Improved |
| replace standard PER with Geo-PER (ablation) | Geo-MoE · 4-task continual | avg N-NBT: 29.6% -> 26.5% | Improved |
| use Geo-PER instead of EWC regularization (ablation) | Geo-MoE · 4-task continual | avg success rate: 38.3% -> 63.3% | Improved |
| add observation residual (point cloud encoder) instead of action residual (ablation) | Geo-MoE · 4-task continual | avg success rate: 3.3% -> 45.8% | Improved |
| add MoE without observation residual (ablation) | action-residual baseline · 4-task continual | avg success rate: 3.3% -> 9.2% | No clear effect |
| add MoE on top of observation residual (ablation) | Geo-MoE · 4-task continual | avg success rate: 45.8% -> 55.8% | Improved |
| transfer from similar pre-trained task (Pick Cube) with 10 trajectories | GeCo-SRT on Stack Cube · Stack Cube | success rate: 26.7% -> 40.0% | Improved |
| transfer from dissimilar pre-trained task (Plug Insert) | GeCo-SRT on Stack Cube · Stack Cube | success rate: 26.7% -> 16.7% | Hurt |
| use continual transfer instead of training from scratch with 20 trajectories | GeCo-SRT on Pick Cube · Pick Cube | success rate: 76.6% with 20 trajectories | Improved |
| use 3 experts instead of 2 or 8 (ablation) | Geo-MoE · appendix sensitivity | avg success rate (%): 2: 60.0, 3: 66.7, 8: 65.0 | Mixed |
| keep only 50% of historical data in replay buffer (ablation) | GeCo-SRT · Pick to Stack Cube | N-NBT: 20.0% -> 23.3% | No clear effect |
Closest papers by meaning
- Overcoming the Sim-to-Real Gap: Leveraging Simulation to Learn to Explore for Real-World RL (NeurIPS 2024 · similarity 0.84)
- Rapidly Adapting Policies to the Real-World via Simulation-Guided Fine-Tuning (ICLR 2025 · similarity 0.84)
- SRSA: Skill Retrieval and Adaptation for Robotic Assembly Tasks (ICLR 2025 · similarity 0.83)
- Generalizable Domain Adaptation for Sim-and-Real Policy Co-Training (NeurIPS 2025 · similarity 0.83)
- DexScale: Automating Data Scaling for Sim2Real Generalizable Robot Control (ICML 2025 · similarity 0.83)
- PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.83)
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.
