AGIBOT's audio-visual model tops Daily-Omni benchmark at 85.21%

AGIBOT announced that its WITA-Omni Preview multimodal foundation model has achieved the highest score on the Daily-Omni audio-visual reasoning benchmark, recording an average accuracy of 85.21 percent. The model outperformed Alibaba's Qwen3.5-Omni-Plus, Google's Gemini 3.1 Pro Preview, and ByteDance's Doubao Seed 2.0 Lite. It achieved the highest scores in audio-visual alignment, comparison, event sequencing, and the benchmark's 30-second and 60-second video subsets, and ranked first or joint first in six of the eight reported evaluation metrics, including tying for first place in inference.

Daily-Omni is a third-party benchmark containing 684 real-world videos and 1,197 multiple-choice questions across six task categories: audio-visual alignment, event sequencing, inference, and reasoning. Unlike benchmarks focused primarily on image-text understanding, Daily-Omni evaluates whether AI models can associate sounds with visible events, understand how situations develop over time, and perform reasoning across multiple data types. AGIBOT states these capabilities are particularly important for embodied AI systems (robots operating in dynamic environments) that must determine who is speaking, associate sounds with actions, follow sequences of events, and decide whether, when, and to whom they should respond.

WITA-Omni differentiates itself through a unified "Thinker-Talker-Actor" architecture. The Thinker functions as the multimodal reasoning engine, processing text, images, audio, and combined audio-visual inputs within a shared representation space. The Talker generates speech in real time based on the Thinker's internal state, while the Actor produces coordinated physical movements and facial expressions. This unified design allows perception, reasoning, speech, movement, and facial expression generation to operate within a shared model state and timeline, enabling robots to continue observing their surroundings while preparing and delivering responses.

The model was trained on tens of millions of hours of multimodal data during continued training, expanding its capabilities beyond text and images to include audio perception and cross-modal reasoning. AGIBOT also developed a human-centric multimodal interaction dataset spanning thousands of hours, which preserves temporal relationships between audio, video, language, body movement, and facial expressions. Training used a three-stage process: supervised fine-tuning (establishing multimodal understanding and coordinated output), on-policy distillation (transferring knowledge from a teacher model with additional contextual information to a student model without it), and reinforcement learning focused on improving interaction decisions including response accuracy, timing, target-person selection, and determining whether a response should be given. The company used Group Relative Policy Optimization (GRPO) to optimize the model's policy.

AGIBOT said WITA-Omni Preview forms the foundation of its Interaction Intelligence platform, being developed alongside its Manipulation Intelligence and Locomotion Intelligence technologies as part of its "Three Intelligences in One" architecture. The Shanghai-based company plans to continue developing the WITA family of multimodal models and integrate the technology into its robotic platforms—including humanoid robots, quadrupeds, dexterous manipulation systems, and commercial cleaning robots—to enable more natural and context-aware interactions between people and embodied AI. In June 2026, AGIBOT announced that its 15,000th robot had rolled off the production line.