Capek 0.5: An Execution-Centric Vision-Language Model for Embodied Intelligence

The Core Problem: Robot Brains Need Multiple Skills at Once

Vision-language models (AI systems that understand both images and text) are becoming the reasoning engine for robots. But here's the catch: robots don't solve a problem once and move on. They act, observe the result, reason about what changed, and decide what to do next. This cycle repeats continuously. Each action reshapes the world, so a robot needs to constantly perceive, think, and verify—and these are fundamentally different mental tasks. Current approaches train each skill separately (spatial understanding here, action planning there), leaving a gap: how do you actually combine them for real robot execution?

Capek 0.5 reorganizes how embodied AI learns by grouping these diverse skills not by dataset or task, but by their functional role in the execution loop. The authors identify four capability families: Spatial Reasoning (where are objects?), Temporal Understanding (how does the scene evolve?), Action Guidance (what should the robot do?), and State Verification (did the action work?).

How It Works: Specialists Then Merge

The method unfolds in two phases. First, each of the four capability families is trained as a specialist model using reinforcement learning with verifiable rewards (rewards that can be objectively checked—e.g., "did the gripper close on the target?"). Each specialist pulls from the same backbone model, so they share foundational vision-language understanding but diverge in their specific training objectives.

Then comes the integration problem: you can't run four separate models on a robot with limited compute. The authors use two techniques to consolidate specialists into one unified checkpoint. Weight-space merging combines the trained parameters from all four specialists. That single merged model is then refined through routed policy-space distillation—a process where different parts of the unified model are taught to route different types of inputs to their appropriate reasoning pathways.

Why This Matters

The authors test Capek 0.5 at two scales (2 billion and 35 billion parameters) across three angles. First, they benchmark against existing datasets and introduce Capek-StateBench, a new benchmark specifically for verifying whether a robot correctly understands the state after an action. Second, they measure capability retention: when specialists merge, do they lose their skills? Results show measurable but acceptable losses. Third, they run closed-loop simulations where robots execute real multi-step tasks—the acid test of whether this design actually works in practice.

The key insight is structural: by organizing training around what execution actually needs rather than what's convenient to dataset-ify, the model learns complementary capabilities that integrate naturally. This isn't just academic reorganization—it reflects how embodied reasoning must work. A robot that can reason about space, predict dynamics, guide its own actions, and catch its own mistakes is closer to robust autonomous operation than one trained on isolated benchmarks.

The method transfers to closed-loop execution, suggesting the execution-centric framing captures something real about what embodied intelligence requires.