Agibot Releases Embodied Interaction Model; Hesai Robot Modules Scale Up

Highlights

• Agibot releases WITA-Omni, with interaction response as fast as about 400 milliseconds
• Hesai robot power modules deliver 23,405 units in the quarter, up 96.8% quarter-on-quarter
• Disney tests AI parking robot on a non-humanoid path

Embodied Intelligence: The Interaction Layer Starts Setting Hard Metrics

Agibot on Nov 10 released WITA-Omni (Silicon Photonics Motion Language All-Modal Large Model), which it calls the industry's first embodied all-modal end-to-end interaction model. Compared with common multimodal models, the difference lies at both the input and output ends: besides audio and video, it adds inputs such as touch and body state, while the output side includes robot actions and facial expressions. Built on a streaming full-duplex architecture, the company gives an interaction response of as fast as about 400 milliseconds and about 600 milliseconds on average, nearly 3 times faster than the industry average. (Source: IT Home)

Another detail worth noting is "wake-word-free" operation: the system combines conversation orientation, video, audio, voiceprint, lip movement, posture, semantics and context to judge whether the current speech needs a response, and the company says it can reduce false triggers amid multi-person conversation or ambient broadcast interference. The model also supports streaming simultaneous interpretation in more than 100 languages and dialects, as well as dozens of emotions and their intensity levels. In July this year, WITA-Omni Preview topped the DailyOmni all-modal understanding leaderboard with a score of 85.21. (Source: IT Home)

Assessment: Interaction latency and false-trigger rates are threshold metrics for whether embodied products can enter real scenarios. Vendors are starting to proactively give verifiable numbers, showing competition is shifting from "can demo" to "can deliver." But these figures are still official claims and need third-party reproduction.

Upstream: Robot Joint Modules Enter Mass Production and Scale Up

Hesai Technology reported revenue of 1.07 billion yuan for the third fiscal quarter of 2026, up 34.3% year-on-year; net profit attributable to shareholders was 80.79 million yuan, up 13.7% year-on-year, marking six consecutive quarters of GAAP profitability. More relevant to the robotics sector is its strategic growth business: revenue of 90.51 million yuan, up 101.4% quarter-on-quarter; robot power modules delivered 23,405 units in the quarter, up 96.8% quarter-on-quarter, monthly capacity rose to 10,000 units, full-body joint modules have entered mass production, and cumulative shipments reached 35,295 units by the end of September, already used in multiple products of AI robotics company Sharpa. (Source: IT Home)

The company expects full-year shipments of robot power modules to reach six-digit scale in 2027. Its core lidar business delivered 729,673 units this quarter, up 65.3% year-on-year; in September it reached a partnership with JD.com to empower its unmanned logistics vehicles and robotics business with 6D full-color lidar. (Source: IT Home)

Assessment: Joint modules moving from "samples" to "10,000 units per month" is an early signal that the embodied intelligence supply chain can scale; but the current base is still small, and whether the six-digit guidance can be met depends on downstream complete-machine shipment pace.

Scenario Side: Non-Humanoid Robots Land First

Disney is working with robotics company Field AI to test an AI-equipped parking guidance robot, which made its debut at the Disney Accelerator demo day. The robot does not use a humanoid design: a four-wheel chassis, orange body with bright yellow accents mimicking the colors of a reflective vest; a large display shows rotating eyes and moving eyelids, inspired by "Cars" and "WALL-E"; a top sign can change direction, and two small robotic arms with red indicator lights are used to guide vehicles into parking spaces. Disney has not announced a deployment timeline or confirmed that it will definitely be put into use. (Source: IT Home)

There is also progress in space: a robot in the ASTRA project developed by Caltech and funded by DARPA assembled separate parts into an antenna truss support structure in orbit, which researchers call the first autonomous structure construction in space. Due to time constraints, only the third part of the truss assembly was completed this time, not a complete antenna; the ground team still needs to use several communication windows of 15 minutes each day to check photos and troubleshoot. (Source: IT Home)

Assessment: These two threads point to the same conclusion — robots are landing first in scenarios with clear task boundaries and greater tolerance for error, rather than directly replacing humanoid general-purpose work.