| name | robo-dopamine-rl |
| title | Robo-Dopamine: General Process Reward Modeling for Robotic Manipulation |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2512.23703 |
| keywords | ["robotics","reinforcement-learning","reward-modeling","multi-view","process-rewards"] |
| description | Overcome reward function design challenges via General Reward Model (GRM) for step-wise progress assessment. Uses multi-view observations for occlusion robustness, hop-based progress normalization, Policy-Invariant Reward Shaping—enabling 95% robot task success within 150 interactions with theoretically-grounded dense rewards. |
Overview
Addresses reward design challenges in robotic manipulation via principled progress modeling.
Core Technique
General Reward Model:
progress = model.predict_progress(observation, task_span)
fused_progress = aggregate_views([view1, view2, view3])
reward = outcome_reward + policy_invariant_shaping(fused_progress)
Performance
- 92.8% task completion accuracy
- 95% success within 150 interactions
- Theoretically grounded reward shaping
References
- Step-wise progress discretization
- Multi-view observation fusion
- Policy-Invariant Reward Shaping