Capek 0.5: New Vision-Language Model for Embodied AI

This research presents Capek 0.5, an advanced vision-language model specifically developed for embodied AI agents. The core idea behind Capek 0.5 is an "execution-centric capability taxonomy," which organizes the agent's skills based on their functional roles throughout an iterative execution process. This contrasts with traditional approaches that often develop capabilities in isolation for specific tasks, leading to integration challenges. Embodied agents require a continuous cycle of perception, reasoning, action, and verification as each action alters the physical state and scene. Capek 0.5 addresses these demands by defining four key capability families: Spatial Reasoning, Temporal Understanding, Action Guidance, and State Verification. Each of these capabilities is initially acquired by a dedicated specialist model through reinforcement learning, utilizing verifiable rewards from a shared backbone. These specialists are then consolidated into a single, unified inference-time model using weight-space merging followed by routed policy-space distillation. The model was instantiated at both 2B and 35B-A3B scales and rigorously evaluated. Results from comprehensive benchmark suites, including a new benchmark for state verification (Capek-StateBench), confirmed that Capek 0.5 significantly improves performance across most matched benchmarks compared to its initialization. It successfully retains all four specialized capabilities within a single checkpoint, with quantified losses, and demonstrates effective transfer to closed-loop embodied task execution in simulated environments.