Aether AI Demonstrates CRIS-0, a Causal Intelligence Robotics System for Adaptive Task Execution

Making coffee and tidying a living room are no longer unfamiliar tasks for today's robots. What is truly difficult is when conditions change mid-execution: the coffee bag is moved by someone, or a hand suddenly reaches in while closing the microwave door. Traditional end-to-end models mainly learn a mapping from visual input to action output, while continuous tasks further require the system to continuously understand and manage task state.

The answer offered by Aether AI's causally driven robotic intelligence system CRIS-0 is to place task state at the center of system design. Here, task state is not a simple description of the current scene, but a set of relevant conditions that determine whether the next action can hold—the position of the coffee bag, whether an object is gripped securely, whether someone is near the door can all change subsequent actions; other irrelevant changes in the scene may not necessarily need to interrupt the task.

The increment of this approach is: the robot must not only know how to complete an action, but also know which conditions the current stage depends on, and update these conditions based on execution results. CRIS-0 puts these judgments into the same operating loop, attempting to give adjustments in continuous tasks a traceable basis.

CRIS-0 uses task state as a shared basis to coordinate the world model, action policy, and verification tools, and on this basis decides whether an action should continue, retry, or fall back. The planner divides the task into observable, verifiable stages, sets completion conditions and reference tools for each stage, and gradually forms a task graph; the task graph records transition relationships between stages and also provides a path for fallback after failure.

On the prediction side, CRIS-0 introduces the company's self-developed Causal World Model, which first predicts how task-relevant states will change, then represents the outcome after action accordingly. Aether AI also adds a module that generates actions to the system, sharing representations with the world model, and the two together form the Policy model.

On the execution side, CRIS-0 uses a unified tool interface to let the planner call rule functions, the policy model, navigation tools, and verifiers, with tool selection depending on which stage the task has reached and what conditions that stage must satisfy. After each stage ends, the verifier checks whether the result meets the preset conditions, and the planner then decides to continue, retry, or fall back.

The team also designed a Structured State Transition Loop, which records successful and failed attempts and incorporates effective recovery paths into the task graph for reuse in similar stages. The intent of this design is to give failure recovery a clear starting point, rather than making it untraceable within an end-to-end model.