tanhuajie2001/Robo-Dopamine-GRM-3B
Robo-Dopamine-GRM-3B is a 3 billion parameter vision-language model developed by tanhuajie2001, designed for high-precision robotic manipulation. It functions as a General Reward Model (GRM) that predicts relative progress or regress hops in robotic tasks by analyzing multi-view images of initial, goal, "BEFORE," and "AFTER" states, conditioned on a task description. This model employs Multi-Perspective Progress Fusion for stable reward signals and is integrated into a Dopamine-RL framework for one-shot adaptation and policy-invariant reward shaping, making it universally compatible with various RL algorithms for accelerating learning in robotics.
Loading preview...
Robo-Dopamine-GRM-3B: General Process Reward Modeling for Robotics
Robo-Dopamine-GRM-3B is a 3 billion parameter vision-language model specifically engineered for high-precision robotic manipulation. Developed by tanhuajie2001, this model introduces a novel approach to reward modeling in robotics, aiming to provide stable and accurate progress signals for reinforcement learning.
Key Capabilities
- General Reward Model (GRM): Acts as a vision-language model that interprets task descriptions and multi-view images (initial, goal, "BEFORE," and "AFTER" states) to predict the relative progress or regress of a robotic task.
- Multi-Perspective Progress Fusion: Combines incremental, forward-anchored, and backward-anchored predictions to generate a robust and fused reward signal, ensuring stability and accuracy.
- One-Shot GRM Adaptation: The Dopamine-RL framework allows the pre-trained GRM to adapt to new tasks using just a single demonstration.
- Policy-Invariant Reward Shaping: Converts the GRM's dense output into an effective reward signal that accelerates learning without altering the optimal policy, ensuring broad compatibility with various RL algorithms.
Good for
- Developing and training robotic manipulation policies that require precise progress feedback.
- Accelerating reinforcement learning in complex robotic tasks through effective reward shaping.
- Researchers and developers working on vision-language models for embodied AI and robotics.