To accomplish long-horizon tasks with defined goal states, we propose ManualVLA, a unified VLA model built upon a MoT architecture. The framework consists of two experts: a planning expert responsible for generating multimodal manuals, and an action expert responsible for predicting precise actions. The planning expert processes human instructions, the current image, and the final goal image to generate intermediate manuals that combine next-step image, positions, and sub-task instructions. We introduce an explicit CoT reasoning process, where each positional indicator serves as a visual prompt embedded into the observation of the action expert. Along with the cross-task shared attention mechanism and the designed attention mask, the generated manual tokens are also used as conditioning signals for action generation, enabling an implicit CoT reasoning process that effectively guides the action expert. For training, ManualVLA adopts a three-stage training strategy that aligns the planning and action experts for effective collaboration.

Action Generation. We design three long-horizon tasks with defined goal states, challenging the model’s procedural reasoning and manipulation capabilities. (1) 2D LEGO Assembly: The task begins with several LEGO bricks of different colors placed on a planar board. Given the final 2D assembled structure as the goal, the model must infer a sequence of intermediate manipulation actions and execute them through coordinated bimanual control. (2) 3D LEGO Assembly: The task extends the 2D LEGO Assembly task to a more challenging 3D setting, where the final configuration transitions from a planar layout to a 3D structure. This upgraded task imposes greater demands on the model’s spatial reasoning abilities. (3) Object Rearrangement: The task begins with several objects of diverse shapes, sizes, and semantics scattered around a box. Given a goal state in which all objects are placed at their designated positions inside the box, the model must progressively generate manipulation actions, alternating control of the left and right arms to prevent collisions.

Compared with the strongest hierarchical baseline, ManualVLA improves the final task completion rate by 15%-30%. While baseline models often succeed in the early stages of a long-horizon pipeline, they typically fail to sustain this performance through the whole sequence. In contrast, ManualVLA mitigates this degradation by decomposing complex tasks into structured subgoal manuals and grounding them into precise actions through a combination of explicit and implicit reasoning, enabling consistent performance throughout the entire task. 

To systematically evaluate the ability of ManualVLA to manipulate on general tasks, retaining the basic capabilities of the VLA model, we conduct experiments on 10 tasks in the RLBench benchmark, which is based on the CoppeliaSim simulator. For each task, we collect 100 demonstration trajectories  through the Open Motion Planning Library. 

ManualVLA obtains an average success rate of 70% across 10 diverse tasks, surpassing the previous SOTA methods π0 and CoT-VLA by margins of 7% and 11%, respectively.