Why Robots Need a “Thinking System”: τ0‑VLA Enables Long‑Term Task Execution

Background and Challenge

Recent WAIC events highlight a shift in embodied intelligence from single‑skill demos to real‑world, multi‑step tasks such as room cleaning, cooking, and warehouse picking. The core bottleneck is that current VLA models are reactive: they map the current observation and language instruction directly to low‑level actions without explicit modeling of task history, progress, or future consequences.

This reactive paradigm leads to three systematic failures in long‑term tasks: error accumulation, step repetition or omission, and goal drift.

τ0‑VLA Architecture

τ0‑VLA introduces a hierarchical “slow‑thinking, fast‑execution” design that decouples high‑level task planning from low‑level motion control. The high‑level policy (slow thinking) performs task decomposition, maintains execution memory, and runs test‑time scaling with a world model to predict future states. The low‑level policy (fast execution) receives the selected sub‑task and generates stable, high‑frequency whole‑body actions using a unified 40‑dimensional action space.

Propose Model : Generates multiple candidate sub‑tasks based on the instruction, current observation, and execution history.

World Model : Predicts the future visual state for each candidate without making value judgments.

Value Model : Scores each predicted future on task completion, safety, and hidden risk.

Reflective Model : Compares candidates, selects the best, and triggers replanning if needed.

This four‑module pipeline implements a sub‑task‑level beam search that evaluates the chain‑of‑physical‑impact of each candidate before execution.

Training Data

τ0‑VLA is trained on 40 115 hours of real‑world robot interaction data, including over 20 000 hours of true‑robot experience across platforms such as AGIBOT G1, ARX AC One, and Franka. The data cover diverse robot morphologies, tasks, and environments, enabling the model to learn universal physical interaction patterns.

Experimental Evaluation

Long‑term tasks were evaluated on real robots (e.g., full‑house cleaning, cooking, milk‑tea preparation, drawer search, desktop organization). On the AGIBOT G1 platform, τ0‑VLA raised the average success rate from 27.5 % to 45.0 % and the average task progress from 80.10 % to 87.85 %, outperforming π0.5 and GR00T N1.7.

In direct‑execution tests without invoking the high‑level planner, τ0‑VLA achieved the highest success rates across four robot platforms, with three tasks reaching 10/10 success, surpassing GR00T N1.7, LingBot‑VLA, and π0.5.

Implications

τ0‑VLA demonstrates a shift from “action intelligence” to “task intelligence” by combining large‑scale real‑world pre‑training, world‑model‑guided test‑time computation, and a closed‑loop learning pipeline (pre‑training → deployment → LWD reinforcement). This architecture shows that high‑level planning and test‑time reasoning are essential for reliable execution of multi‑step, long‑duration tasks in open environments.

The work suggests that integrating massive real‑world data with hierarchical planning and test‑time computation is a viable path toward scalable deployment of embodied AI systems.