Chinese humanoid-robot maker AgiBot has unveiled an interaction model designed to let its machines converse and react the way people do in noisy, crowded rooms. The company announced WITA-Omni on the evening of November 10 and demonstrated it on a mass-produced Expedition A3-Ultra robot, which held a roughly ten-minute live exchange with an executive in what the company called an extreme stress test of the system.
What distinguishes WITA-Omni from earlier robot assistants is the range of senses it takes in. Most existing systems rely on cameras and microphones, running understanding, dialogue and motion as separate stages. The new model adds touch and proprioception, awareness of the robot’s own body position, to its inputs, and produces robot movements and facial expressions alongside speech. AgiBot argues that physical contact carries meaning a camera alone misses: the pressure of a handshake or the direction of an outstretched hand can change what a person intends.
The system is built on a streaming, full-duplex architecture, meaning the robot listens, thinks, answers and moves at the same time rather than completing one stage before starting the next. It can nod or hold eye contact while a person is still speaking, and it judges in real time whether a pause means the speaker has finished. AgiBot said response latency can be as low as 400 milliseconds and averages around 600, roughly three times faster than the industry norm. A full-modal rejection system combines where a person is facing, voiceprints, lip movement, posture and context to decide whether speech is aimed at the robot, so no wake word is needed and background announcements are less likely to trigger a reply.
The model handles more than 100 languages and dialects without manual switching and can perform streaming, two-way simultaneous interpretation. On its face screen it renders dozens of emotions, from a faint smile to hearty laughter, and it can draw or visualize complex information in real time when words alone fall short. Perhaps most important for real deployment, the large-parameter model runs on the robot’s own onboard computing rather than in the cloud, which AgiBot said protects privacy, allows offline interaction and removes network round trips.
The company said WITA-Omni has reached first place on four embodied-interaction benchmarks, covering omni-modal understanding, temporal localization in video, audio understanding and embodied spatial reasoning. Robots carrying the model are already working in public settings, handling knowledge presentations and visitor guidance at entertainment venues and theme-park tours. If the approach proves durable in factories, hotels and malls, where networks are unreliable and several people talk at once, the ability to keep working entirely on-device may matter more than any single benchmark score.

Dgcity