srpo-vla-policy

Train VLA models for robotic manipulation by using the model's own successful trajectories as self-reference for reward—enable progress-based feedback for failed attempts without external rewards or demonstrations.

Independent GitHub snapshot · Automated classification · Original documentation governs