Physical AI News (Japan)

Physical AI, robotics and compute — reported from Japan.

日本語
· Product launches· 日本語で読む

AgiBot unveils omni-modal dialogue model WITA-Omni, with fastest response of about 400 milliseconds

AgiBot announced its dialogue model for robots, WITA-Omni, on Nov. 16, 2026. It takes audio and video, as well as tactile input and robot state, and outputs movements and facial expressions. The company says it runs on the Expedition A3-Ultra and responds in as little as about 400 milliseconds.

Key facts

  • AgiBot announced its omni-modal dialogue model for robots, WITA-Omni, on Nov. 16, 2026.
  • Inputs include audio and video, as well as tactile information and robot state; outputs include robot movements and facial expressions.
  • The company says response time is as fast as about 400 milliseconds and averages about 600 milliseconds, about three times faster than the industry average.
  • The company described onboard operation on the Expedition A3-Ultra and a configuration combining the Interaction Engine and Background Agent.
  • The company says WITA-Omni supports more than 100 languages and dialects.

What happened

AgiBot announced its omni-modal dialogue model for robots, WITA-Omni, on Nov. 16, 2026. In addition to audio and video, it takes tactile input and the robot's state, and outputs movements and facial expressions. Response time is as fast as about 400 milliseconds and averages about 600 milliseconds, which the company says is about three times faster than the industry average.

Why it matters

WITA-Omni was announced as a model that takes tactile input and robot state and outputs movements and facial expressions, aiming to provide an interface between people and robots that goes beyond audio responses. AgiBot cites onboard operation as an advantage for privacy, dialogue in environments without connectivity and real-time processing. It says, however, that complex tasks use a cloud-based Background Agent, so the conditions for use need to be assessed for each type of processing.

Analysis

WITA-Omni's focus is on linking a robot's recognition of tactile input and its own state to the generation of movements and facial expressions, rather than limiting dialogue to audio input and output. AgiBot described training using data from the robot itself and onboard operation on the Expedition A3-Ultra, outlining an approach that adapts a dialogue model to a robot's physical capabilities. However, the response-speed figures and ranking claims are based on company announcements and cannot be equated with independent evaluations on actual hardware. In relation to our previous coverage, AgiBot unveils embodied-agent framework “Genie Master,” tests autonomous improvement in a real store covered feedback on execution results, while AgiBot tests six Genie G2 robots operating 24 hours a day at a convenience store in Shanghai reported on work and record-keeping in a real store. This announcement gives more concrete form to a layer for interaction that receives human conversation and connects it to responses and gestures, separate from those task-execution systems. The announcement does not explain how these systems connect through shared data infrastructure or training processes. An important aspect of the technical design is the separation of the Interaction Engine, which handles immediate responses, from the Background Agent, which handles planning and longer reasoning, while allowing them to share context. Robot dialogue requires both fast responses and the ability to handle complex requests, so the division of processing appears intended to achieve both. The company says the model runs onboard but that complex tasks are processed in the cloud. Which inputs and processing remain onboard will therefore be a key consideration for data management and use in areas with unreliable connectivity. Further details are needed on the conditions used to measure the fastest response time of about 400 milliseconds, the scope of the average, and response quality with inputs that include tactile information and robot state. The announcement also gives no specific figures for model size, onboard computing resources, the range of tasks requiring a cloud connection, safety in public settings or the scale of deployment beyond tourist guidance. In operational decisions, assessments of misrecognition and consistency with robot movements in real-world environments will matter, alongside dialogue benchmark results.

Details

According to the company, WITA-Omni is a large model that runs onboard the Expedition A3-Ultra. It uses streaming full-duplex processing to generate audio, facial expressions and movements while listening to input. The system combines an Interaction Engine for basic responses with a Background Agent for reasoning, planning and tool use, among other functions; the latter handles complex tasks in the cloud. It supports more than 100 languages, including dialects, and the company says it ranked first in the DailyOmni, ActivityNet-TimeLens, MMAU and EmbSpatialBench evaluation rankings. The company said its training process includes omni-modal pretraining, behavior tuning using data from the robot itself, and dialogue-policy reinforcement. It said robots equipped with WITA-Omni have explained information and guided visitors in some entertainment and tourism settings, such as tourist information.

Sources

Translated from the Japanese original: /news/satoru-moto-soushin-zen-moodaru-taiwa-moderu-wita-omni-wo-ha-1236f8. The Japanese version is authoritative if they differ.