A Chinese AI company has built a multimodal model that outperforms Google's Gemini and Alibaba's systems at understanding audio and visual information simultaneously, achieving 85.21% accuracy on a real-world reasoning benchmark. AGIBOT's WITA-Omni Preview ranked first on the Daily-Omni benchmark, a test designed to measure whether AI systems can connect sounds with visible events, follow sequences of actions, and reason across both modalities at once.

The achievement signals a shift in how embodied AI, or robots designed to interact with the physical world, are being built. Rather than bolting together separate modules for perception, speech, and movement, AGIBOT designed WITA-Omni as a unified system where all these functions share the same internal understanding of what's happening around them.

Why Can't Robots Just Use Separate Hearing and Vision Systems?

Traditional robotic systems treat perception, dialogue, and action as separate tasks. A robot might use one AI model to understand what it sees, another to process speech, and a third to decide how to move. This approach creates timing problems. By the time the robot finishes analyzing a sound, the situation may have changed. A person might have already moved, or the conversation might have shifted to a new topic.

WITA-Omni takes a different approach. The model maintains a shared timeline and state across all modalities, meaning it continues observing its surroundings and receiving audio input while simultaneously preparing its response and generating corresponding speech, movements, and facial expressions. This unified architecture is called the "Thinker-Talker-Actor" framework.

How Does the Thinker-Talker-Actor Architecture Actually Work?

• Thinker Component: Serves as the multimodal reasoning core, mapping text, images, audio, and audio-visual inputs into a shared representation space and performing cross-modal reasoning and high-level interaction decision-making.
• Talker Component: Generates natural speech in real time, conditioned on the Thinker's internal states, allowing the robot to respond conversationally without delay.
• Actor Component: Uses dedicated motion and facial-expression heads to generate physical movements and expressions alongside speech, extending the model from digital interaction to embodied output.

This three-part design allows a robot to coordinate what it perceives with how it responds. Rather than passing information sequentially through independent modules, the architecture maintains a shared model state and timeline across modalities.

What Makes WITA-Omni's Performance Stand Out on Daily-Omni?

The Daily-Omni benchmark is not a typical AI test. It comprises 684 real-world videos and 1,197 multiple-choice questions across six task categories, including audio-visual alignment, event sequencing, inference, and reasoning. Unlike benchmarks that focus primarily on image-text alignment or short-form visual understanding, Daily-Omni tests whether a model can connect sounds with visible events, follow how situations unfold over time, and reason across audio and visual information.

WITA-Omni Preview achieved an average accuracy of 85.21%, outperforming Qwen3.5-Omni-Plus, Gemini 3.1 Pro Preview, and Doubao Seed 2.0 Lite. The model recorded the highest scores in audio-visual alignment, comparison, event sequencing, and the benchmark's 30-second and 60-second video subsets. It also tied for first place in inference, ranking first or joint first in six of the eight metrics reported on the leaderboard.

These capabilities are particularly relevant to embodied AI in dynamic, multi-person environments. A robot must be able to identify who is speaking, associate sounds with visible events, follow sequences of actions, and determine whether, when, and to whom it should respond.

How Was WITA-Omni Trained to Achieve This Level of Performance?

WITA-Omni's performance is driven not simply by model scale, but by a tiered data framework and training methods developed specifically for multimodal interaction in physical environments. During continued training, WITA-Omni was trained on multimodal data at the scale of tens of millions of hours, extending its text and visual capabilities to support audio perception and joint reasoning across sound and visual information.

However, publicly available audio-visual data often lacks important signals found in real interactions, including conversational turn-taking, response timing, body movements, and facial expressions. To address this gap, AGIBOT developed a large-scale, human-centric multimodal interaction dataset spanning thousands of hours. The dataset preserves the temporal relationships among audio, video, language, body movement, and facial expression, helping the model learn how real-world interactions unfold over time.

WITA-Omni was subsequently trained using a three-stage process comprising supervised fine-tuning, on-policy distillation, and reinforcement learning.

• Supervised Fine-Tuning (SFT): Establishes the model's core capabilities in multimodal understanding, interaction decision-making, and coordinated multi-stream generation.
• On-Policy Distillation (OPD): Uses a teacher version of the model with access to additional contextual information, such as relevant time intervals and target identities, to produce higher-quality outputs. These capabilities are then distilled into a student version that operates without the additional information, improving audio-visual reasoning while preserving the model's response style.
• Reinforcement Learning (RL): Further improves interaction decision-making, including response accuracy, timing, target-person selection, and whether the model should wait or respond. Group Relative Policy Optimization (GRPO) is used to optimize the model's policy.

Through this process, the model learns not only to generate accurate responses, but also to determine whether a response is appropriate, when it should respond, and to whom the response should be directed.

What Does This Mean for the Future of Robot Interaction?

WITA-Omni Preview's performance on Daily-Omni demonstrates its capabilities in audio-visual temporal understanding and cross-modal reasoning, two areas that are important to human-robot interaction in dynamic physical environments. WITA-Omni is designed to move embodied interaction away from a sequence of loosely connected modules and toward a continuous process in which perception, decision-making, and expression are coordinated within a shared model state and timeline.

AGIBOT announced in June 2026 that its 15,000th robot had rolled off the production line, indicating that the company is scaling production of embodied systems. WITA-Omni is designed to serve as the foundation of AGIBOT's Interaction Intelligence, which will continue to evolve alongside the company's Manipulation Intelligence and Locomotion Intelligence under its "Three Intelligences in One" architecture.

The company stated that it will continue advancing the WITA family of multimodal models and integrating these capabilities into its robotic platforms, with the goal of enabling more natural, timely, and context-aware interaction between people and embodied AI.