Research map / Reinforcement learning / RL from feedback
General Process Reward Modeling for Robotic Reinforcement Learning
Huajie Tan, Sixiang Chen, Yijie Xu, Zixiao Wang, Cheng Chi, Yuheng Ji, Yaoxu Lyu, Zhongxia Zhao, Xiansheng Chen, Peterson Co, Shaoxuan Xie, Guocai Yao and 3 more
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 policy-invariant reward shaping instead of raw progress increment reward (ablation) | Dopamine-RL · real-world 3 tasks | average success rate (%): 41.3 -> 85.0 | Improved |
| add one-shot GRM adaptation from a single demonstration (ablation) | Dopamine-RL · real-world 3 tasks | average success rate (%): 63.2 -> 85.0 | Improved |
| add multi-perspective progress fusion over incremental-only (ablation) | Dopamine-RL · real-world 3 tasks | average success rate (%): 70.0 -> 85.0 | Improved |
| add fusion over forward-anchored-only (ablation) | Dopamine-RL · real-world 3 tasks | average success rate (%): 65.7 -> 85.0 | Improved |
| add fusion over backward-anchored-only (ablation) | Dopamine-RL · real-world 3 tasks | average success rate (%): 62.5 -> 85.0 | Improved |
| use multi-view instead of single-view inputs for GRM-8B (ablation) | GRM-8B · task completion classification | accuracy (%): 83.9% -> 92.8% | Improved |
| use multi-view instead of single-view for GRM-3B (ablation) | GRM-3B · VOC average, dense sampling | Value-Order Correlation: 0.87 -> 0.93 | Improved |
| scale GRM from 3B to 8B (single-view) (ablation) | GRM · VOC average, sparse sampling | Value-Order Correlation: 0.91 -> 0.92 | No clear effect |
| add dense rewards from GRM via Dopamine-RL vs sparse-reward RL | RL policy · Simulation 10 tasks | success rate (%): 79.9 -> 81.0 | Improved |
| add dense rewards from GRM vs sparse-reward RL | real-world RL · Real-world 8 tasks | success rate (%): 68.0 -> 95.2 | Improved |
| add Dopamine-RL to improve OOD generalization over BC | policy · Insert Square, Circuit, Cap Pen OOD | avg relative drop (%): 57.1 -> 19.3 | Improved |
| denser temporal sampling for baseline reward models | VLAC-2B and GVL · VOC | Value-Order Correlation | No clear effect |
Closest papers by meaning
- Progress Reward Model for Reinforcement Learning via Large Language Models (NeurIPS 2025 · similarity 0.90)
- Efficient Language-instructed Skill Acquisition via Reward-Policy Co-Evolution (AAAI 2025 · similarity 0.89)
- Learning Process Rewards via Visitation Matching for Efficient RL (NeurIPS 2026 · similarity 0.89)
- Text2Reward: Reward Shaping with Language Models for Reinforcement Learning (ICLR 2024 · similarity 0.89)
- R*: Efficient Reward Design via Reward Structure Evolution and Parameter Alignment Optimization with Large Language Models (ICML 2025 · similarity 0.88)
- SOLE-R1: Video-Language Reasoning as the Sole Reward for On-Robot Reinforcement Learning (NeurIPS 2026 · similarity 0.88)
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.
