InternRobotics/VLAC

VLAC (Vision-Language-Action-Critic) addresses a critical bottleneck in robotic foundation models: the transition from pure imitation learning (predicting the next action token) to value-based decision making. By adding a Critic to the VLA architecture, it allows for offline RL and potential online fine-tuning, which is essential for robustness in real-world environments. With 284 stars, it has respectable traction for a specialized robotics repo, though the low fork count (9) suggests it is currently viewed more as an academic reference than a framework being actively extended by the community. It faces extreme frontier risk because the primary developers of VLA models (Google DeepMind with RT-2/RT-X and OpenAI-backed startups like Physical Intelligence) are moving aggressively toward unified 'World Models' that incorporate similar reward-critique mechanisms. The defensibility is moderate; while the 'InternRobotics' (Shanghai AI Lab/SenseTime ecosystem) backing provides high-quality research pedigree, the project lacks the data gravity or proprietary hardware integration needed for a long-term moat against platform giants who can train larger models on more diverse datasets.