What is Thinker
Thinker is an open-source embodied intelligence vision-language large model from UBTECH, purpose-built for robotics scenarios. The 4B-parameter model ranked first globally across 9 authoritative benchmarks. Its core capabilities span task planning, spatial understanding, temporal reasoning, and visual grounding, effectively addressing robots' dilemma of "knowing what to do but not grabbing accurately." The model was trained on 10 million high-quality data points refined from 2 billion raw data points, using an automated annotation system that keeps human involvement below 1%. The model has already supported Walker S2 in achieving a 99.99% task accuracy rate in industrial scenarios, promoting the inclusive development of embodied intelligence technology.
Thinker's Main Features
- Task Planning: Thinker can understand complex human instructions, combine historical state memory, predict future state changes of the robot, and decompose long-horizon tasks into executable sequences of subtasks.
- Spatial Understanding: Thinker establishes an egocentric coordinate system, using the camera as the origin to define all spatial relationships, enabling the robot to accurately perceive the position and orientation of objects in three-dimensional space.
- Temporal Understanding: Thinker can extract key information from video history, combine past events with current instructions, accurately assess the current state, and make reasonable temporal decisions.
- Visual Grounding: Thinker can describe object positions using bounding boxes and precise point coordinates, providing accurate spatial guidance for the robot's grasping operations and interactions.
Thinker's Technical Principles
- Data Construction: Thinker has built a complete pipeline from raw data to high-quality training data. Facing 2 billion noisy, hard-to-align raw data points, it performs broad screening through customized rules, uses large models for multi-dimensional quality scoring, and refines 10 million high-quality data points. At the same time, it adopts an automated annotation system of "large model-assisted annotation plus multi-model cross-validation," keeping human involvement below 1%, reducing annotation costs by 99% while improving efficiency by more than a hundredfold.
- Model Architecture Design: Thinker adopts a classic vision-language model architecture, comprising four core modules: a text tokenizer, a visual encoder, a multi-layer perceptron alignment layer, and a language model backbone. It achieves unified representation of vision, language, and time, enabling the model to accurately capture visual details, understand task instructions, and perform cross-modal reasoning.
- Training Strategy: Thinker adopts a two-stage training method. In the first stage, it is fine-tuned on general datasets, spatial understanding datasets, and large-scale planning datasets to establish basic perception and reasoning capabilities, while introducing the last frame of the video as auxiliary input to enhance video understanding. In the second stage, it undergoes supervised fine-tuning on industrial task datasets, adapting the model to sequence dependencies, diverse object layouts, and feedback correction, ultimately generating planning solutions executable in real industrial scenarios.
- Key Innovations: Targeting the pain points of robot perspective confusion and video information omission, Thinker proposes a simple and effective method of jointly inputting keyframes and the full video during video understanding training, significantly enhancing the model's temporal understanding capability. At the same time, through high-quality data screening and task-oriented sampling, it achieves performance surpassing models with over 10B parameters at a scale of only 4B parameters.
Thinker's Project Address
- GitHub Repository：https://github.com/UBTECH-Robot/Thinker
- HuggingFace Model Library：https://huggingface.co/UBTECH-Robotics/Thinker-4B
- arXiv Technical Paper：https://arxiv.org/pdf/2601.21199
Thinker's Application Scenarios
- Industrial Intelligent Manufacturing: Thinker can drive humanoid robots to complete tasks such as box handling and workpiece sorting on factory production lines. Walker S2 has achieved a 99.99% operation accuracy rate, effectively solving the problem of insufficient flexibility in traditional automation equipment.
- Warehousing and Logistics Operations: Thinker supports robots in performing cargo recognition, path planning, and precise grasping in dynamic warehouse environments, adapting to logistics needs with diverse SKUs and high-frequency changes.
- Commercial Service Scenarios: Thinker empowers robots to provide guidance, explanation, and interactive services in public places such as shopping malls and exhibition halls, achieving natural human-robot interaction through visual-language understanding.
- Complex Operation Tasks: Thinker enables robots to perform operations that require long-horizon planning and fine spatial perception, such as equipment inspection, component assembly, and multi-step experimental procedures.
- Swarm Intelligence Collaboration: As a cognitive foundation, Thinker supports UBTECH's swarm brain network and collaborative intelligent agent Co-Agent, enabling task allocation, collaborative decision-making, and autonomous evolution among multiple robots.