Research map / Vision-language and multimodal / Robot action and navigation
SRPO: Self-Referential Policy Optimization for Vision-Language-Action Models
Senyu Fei, Siyin Wang, Li Ji, Ao Li, Shiduo Zhang, Liming Liu, Jinlong Hou, Jingjing Gong, Xianzhong Zhao, Xipeng Qiu
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 online SRPO post-training to one-shot SFT policy | OpenVLA*-One · LIBERO average | success rate (%): 48.9 -> 99.2 | Improved |
| use online SRPO instead of offline SRPO (ablation) | OpenVLA*-One · LIBERO average | success rate (%): 92.5 -> 99.2 | Improved |
| add self-referential progress-wise rewards instead of sparse outcome rewards (vs RLinf/SimpleVLA-RL) | GRPO-style VLA RL · LIBERO average | success rate (%): 98.0 -> 99.2 | Improved |
| replace latent world representation reward with pixel-level reward (ablation) | SRPO · progress reward benchmark | Spearman correlation: 0.998 -> 0.125 | Hurt |
| replace V-JEPA 2 latent world representation with ImageBind embeddings (ablation) | SRPO · RL training success rate | success rate: plateaus around 85% | Hurt |
| replace latent world representation with ImageBind on reward benchmark (ablation) | SRPO · progress reward benchmark | SMD: 188.799 -> 18.111 | Hurt |
| add SRPO instead of GRPO outcome-only reward | VLA RL · LIBERO-Long and LIBERO-Object training curves | training efficiency | Improved |
| apply SRPO online RL on one-shot SFT policy | OpenVLA*-One · LIBERO-Plus (zero-shot perturbations) | total success rate (%): 19.4 -> 59.6 | Improved |
| apply SRPO on one-shot SFT policy, compared to OpenVLA-OFT zero-shot | OpenVLA*-One · LIBERO-Plus total (zero-shot) | success rate (%): 69.6 vs 59.6 | Hurt |
| train SRPO in LIBERO-Plus perturbed environments | OpenVLA*-One · LIBERO-Plus total (+Data) | success rate (%): 30.7 -> 82.1 | Improved |
| apply progress reward shaping offline to real-world pi0 and pi0-FAST | pi0 / pi0-FAST · five real-world X-ARM 7 tasks | average success rate: +66.8% and +86.7% | Improved |
Closest papers by meaning
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- RePO-VLA: Recovery-Driven Policy Optimization for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- SOLE-R1: Video-Language Reasoning as the Sole Reward for On-Robot Reinforcement Learning (NeurIPS 2026 · similarity 0.90)
- LaST-R1: Reinforcing Robotic Manipulation via Adaptive Physical Latent Reasoning (NeurIPS 2026 · similarity 0.90)
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.
