τ0-VLA Released: Enabling Robots to 'Think Before Acting' for Long-Horizon Tasks

At WAIC 2026, the focus of embodied intelligence shifted from single-skill demos to complex task completion in real-world scenarios. However, current mainstream VLA models are essentially reactive decision-making paradigms that directly map current observations and language instructions to low-level actions, lacking explicit modeling of task history, execution progress, and consequences. This 'look-and-act' approach performs adequately in short-horizon tasks lasting seconds but reveals three systemic flaws in long-horizon tasks lasting minutes and involving dozens of steps: error accumulation (e.g., a 1mm gripper offset or 0.5-degree chassis angle deviation amplifies step by step), step repetition and omission (e.g., picking up a cup that has already been collected), and goal drift (attention captured by local details, behavior gradually deviating from the initial instruction).

Real-world tasks like 'tidy up the room' or 'prepare a meal' often involve dozens of consecutive steps. Robots need not only precise actions but also continuous planning, execution, and error correction around long-term goals. This demands a shift from 'teaching a robot to perform an action' to 'enabling a robot to complete a task,' and τ0-VLA is a new approach to address this challenge.

Inspired by the dual-system theory of human brain science, τ0-VLA splits the robot system into a high-level planning system (slow thinking) and a low-level action system (fast execution). The high-level system understands user goals, decomposes complex tasks into sub-tasks, and tracks execution status via task memory, answering 'what to do next'; the low-level system converts selected sub-tasks into stable, real-time whole-body actions. This decoupling avoids the dilemma of having a single model handle both long-term planning and real-time control, allocating expensive test-time computation to the high level while preserving high-frequency, low-latency action control for the low level.

The high-level system consists of four modules that collaboratively implement a sub-task-level beam search mechanism: a Propose Model generates multiple candidate next sub-tasks; a World Model predicts the environment state after each candidate execution using an image editing model, outputting 'imagined future frames'; a Value Model scores the predicted outcomes across multiple dimensions, assessing task completion, safety, and implicit risks; and a Reflective Model compares all candidates and outputs the final decision, triggering a reselection if it conflicts with overall progress. This is the first time test-time computation has been precisely directed toward forward evaluation of 'chain physical impacts' in the real physical world.

The low-level system constructs a unified 40-dimensional whole-body action space, using masks to disable unused degrees of freedom, enabling cross-embodiment adaptation across various robot morphologies such as single-arm, dual-arm, fixed-base, and wheeled humanoid. During execution, continuous actions are generated in a closed-loop manner via visual feedback, allowing the robot to adjust behavior based on real-time environmental changes, improving stability against real-world perturbations.

τ0-VLA introduces world model-guided test-time computation into embodied decision-making. The world model does not aim to generate photorealistic future scenes but answers 'what changes will occur in the future world if a certain action is executed.' When facing a task, the robot does not immediately choose an action but first predicts future states for different plans. Test-time computation gives the robot a 'thinking budget' capability: simple problems maintain efficiency, while complex problems invest more reasoning resources. Coupled with a deep thinking dynamic routing mechanism, the model only initiates deep thinking at sub-task nodes judged as difficult, outputting fast decisions otherwise, balancing decision quality and reasoning efficiency.

This mechanism breaks the limitation of traditional LLMs that simply repeat sampling and select the best in text space, marking the first time test-time computation is used to evaluate chain effects in the real physical world. Experiments show that after introducing the hierarchical planning system, τ0-VLA's average success rate on the AGIBOT G1 platform increases from 27.5% to 45.0%, and average task progress from 80.10% to 87.85%, significantly outperforming mainstream VLA models like π0.5 and GR00T N1.7 in long-horizon tasks.

τ0-VLA is pre-trained on 40,115 hours of real-world robot interaction data, including over 20,000 hours of real robot data, covering multiple robot platforms such as AGIBOT G1, ARX AC One, and Franka, as well as UMI data and several public datasets. This is the largest real-robot data pre-training in the embodied intelligence field to date, encompassing various morphologies including fixed manipulators, mobile robots, and dual-arm robots, forming a multi-robot, multi-task, multi-scenario physical world experience repository.

Training employs a collaborative approach between the high-level planning system and the low-level action system: the low-level policy learns action execution capabilities from robot demonstration data; the high-level planner learns task decomposition and prediction using task decomposition annotations, execution memory, and world state change information; additionally, multimodal joint training data is incorporated to maintain visual-language understanding. This ensures that τ0-VLA trains not just an action model but a robot intelligence system oriented toward real-world tasks.

τ0-VLA is not an isolated work. The team previously released LWD (a post-training reinforcement learning system for the real world) and τ0-WM (a pre-trained embodied world model). Together, these three works form a 'large-scale pre-training + large-scale post-training' pipeline: pre-training provides general capabilities, the world model rehearses action consequences, τ0-VLA completes complex tasks, and data generated from real deployment flows back into training via LWD, creating a continuously evolving data loop. This pipeline aims to break the limitations of static imitation learning, allowing robots to continue learning, correcting, and improving after entering the real world.

From an industry perspective, τ0-VLA demonstrates a system architecture moving from 'action intelligence' to 'task intelligence.' Early competition focused on hardware structure and motion control; the VLA stage shifted to instruction understanding and action execution; τ0-VLA's 'reason first, act later' paradigm moves robots from passive execution to active planning. Combining the generalization capabilities of large models with the reinforcement learning data flywheel of the real physical world is considered the essential path to large-scale deployment of embodied intelligence.