WCM: World-Cognition Model for Generalizable Human–Robot Interaction

Language agents are capable, conversational partners in software, yet robots still cannot truly interact with the people they serve. The two paradigms that dominate embodied control (end-to-end vision-language-action policies and world-model planners) take an instruction once and rarely talk back, so a person cannot easily tell what the robot understood, redirect it mid-task, or calibrate trust. We argue the bottleneck is less the language interface than representational alignment: a robot tends to interact better when it reasons from a human-centric, object-and-affordance view of the world, compatible with the person's own. We present the World-Cognition Model (WCM), built on the SLAK harness: four decoupled, white-box layers (Sensing/CyberView, Logic, Action, and Knowledge) coordinated by an asynchronous runtime that lets the robot think, talk, and explain while it moves rather than freezing to reason. WCM is designed to support transfer across objects, phrasings, and task compositions by reusing explicit reasoning over grounded objects, affordances, and action primitives. WCM further improves from its own operation: a teacher ensemble turns the robot's successful and failed rollouts into refined chains of thought distilled into a small edge model. In preliminary live trials on a low-cost mobile manipulator, WCM reaches a 73.8% average success rate across nine dialogue-driven interaction tasks.

The World-Cognition Model (WCM) is a reasoning-first embodied agent. Unlike vision–language–action policies trained on human teleoperation, WCM improves from its own experience and from interaction with the people it serves, through three components:

The SLAK architecture. WCM is built from four decoupled, white-box layers—Sensing, Logic, Action, and Knowledge. The Logic layer is a robotics-tuned reasoning model that reads an instruction as intent rather than keywords and follows a plan–reason–validate loop, committing to an explicit expected outcome it can state aloud—so a misunderstanding is caught before it becomes a wrong action. The Sensing layer, CyberView, fuses RGB, depth, audio, and proprioception into a single labeled 3D scene (Fig. 2) whose defining product is the interactable point: the precise, metric point the arm acts on (a bin at its lid, a drawer at its handle), turning vague deictic language such as “the bin” or “that one” into a place the robot can reach. An asynchronous runtime splits control into concurrent reasoner, executor, and result loops, so the robot can think, talk, and explain while it moves rather than freezing to reason.

Self-improving CoT distillation. Every rollout logs the full chain of reasoning behind each decision, and both successful and failed traces are kept. The supervised unit is the per-step reasoning—not motor actions, and not whole tasks—so a single task yields 20–40 reusable (context, reasoning) pairs. A teacher ensemble refines these into cleaner chains of thought, distilled into the edge model via LoRA. The gains generalize rather than overfit: held-out skills improve because what transfers is reusable reasoning about handles, affordances, and action sequencing rather than memorized task scripts. Across the nine interaction tasks, success rises from roughly 20% with no fine-tuning to about 74% after training.

Human-in-the-loop skill acquisition. For tasks too novel or long-horizon for the autonomous loop, a person places the robot in a teaching mode by voice and walks it through the task one step at a time (Fig. 1). The entire human–robot interaction—what the person says, what the robot does, and the person's feedback—is recorded, up-weighted, and folded back into distillation: reused immediately from working memory and later internalized into the model's own competence.

We present WCM, a reasoning-first embodied system built on the SLAK architecture and an asynchronous runtime, that achieves interaction through representational alignment—a grounded view of objects, actions, and intent shared between robot and person. It points toward generalizable human–robot interaction: robots that do not merely execute commands, but reason with people through a shared representation of the world.

We thank L.T. and Andrew Hoang for their contributions to the system implementation and experiments. This work was funded by Cyberbrain, Inc. as part of its internal R&D program; the system, datasets, and related IP are owned by Cyberbrain, Inc.