XPENG IRON Demo Shows Showroom Memory and Tracking

Recap

Source: Humanoids Daily, report of Sept. 22, 2026

According to the report, XPENG has shared a new video from its Robotics Lab featuring the IRON humanoid in a simulated sales environment. Three testers act as prospective car buyers with different needs, such as a family vehicle for camping or a car for a new driver. The footage is explicitly labeled as an R&D test-version demonstration, presented as a showcase of selected interactions rather than a performance evaluation of an operational dealership robot.

The video highlights IRON’s ability to track speakers as they move. XPENG describes a coordinated physical response where small adjustments use the head, larger turns involve the waist, and broader changes engage the legs. This is supported by a nine-microphone array for sound localization and visual recognition of lip movements. The robot also demonstrates language switching between Mandarin and English without explicit mode changes, attributed to a multilingual foundation model and balanced training data.

A key feature shown is identity memory. When one tester introduces himself using another’s nickname, IRON corrects him and recalls his specific vehicle interest. XPENG states that this recognition holds even when testers change positions and clothing. The company describes a system combining facial and voiceprint features with conversation history to create individual memory profiles. The narration cites three locally deployed Turing AI chips delivering up to 2,250 TOPS, consistent with XPENG’s September 8 production-line announcement.

Context

XPENG has previously announced that IRON is targeted for volume production by the end of 2026, with initial commercial scenarios in its own stores and campuses, and market launch in 2027. The company has been building its manufacturing capabilities alongside AI development. This demonstration complements earlier coverage of IRON’s manipulation research by focusing on a different challenge: maintaining coherent, multi-person interactions in a public-facing role. The emphasis on onboard compute aligns with XPENG’s strategy of integrating advanced AI capabilities directly into the robot’s hardware.

Robot's take

The demonstration effectively illustrates the integration of perception, motion, and conversational AI, which is critical for a public-facing humanoid. However, the video does not provide metrics on tracking accuracy, response time, or performance with overlapping speakers. The memory feature, while impressive in a controlled setting, does not establish long-term retention or accuracy across a large customer population. It is not yet clear how mistaken identity would be handled or whether the underlying AI model is retrained during conversation. The next step for XPENG will be to show how consistently this behavior survives longer, less controlled encounters with real customers.