XPENG Robotics, part of the Chinese electric-car maker XPENG, introduced XPACE on Sept. 15, 2026, an artificial intelligence system that teaches the company's IRON humanoid robot partly from videos of people. The usual way to teach a robot, recording people as they steer it through each task, is costly to repeat across many settings. In the company's video, IRON moves bread from a toaster to a plate on its own, a task it never practiced with a human operator; XPENG says the skill carried over from videos of people.
XPACE learns from a huge library of video, from footage filmed through people's own eyes to recordings of IRON guided by human operators. One part of the system decides what the robot should do next, while another predicts what the scene will look like after a given move, much as a person pictures a pour before tipping the jug. Because one shared model learns from both human and robot footage, XPENG says, a skill seen in human video can carry over to the robot.
XPENG also turned that ability to predict into a practice space. The model pictured IRON drifting off course and recovering, and the robot then trained on the best of those imagined recoveries. XPENG reports that water pouring went from working half the time to 19 times in 20. The same setup lets engineers watch the robot's choices play out in imagination before a real test, where it caught mistakes such as reaching for a red fruit when asked for a green one.
XPENG also has plans for IRON itself. The company says its robots are scheduled to enter mass production by the end of this year, with the first ones going to work in its own stores and campuses and sales in China and abroad planned for 2027. Looking further ahead, the XPACE team describes a robot that chooses what to learn next, picking up ideas from people, testing them in imagination and checking them in the real world. Getting there, the team writes, will take models that know when they are unsure and stay reliable through longer, hands-on tasks.
XPENG's public code page lists no XPACE release, and its results come from the company's own trials. The team also notes that the human videos help only with skills that appear somewhere in its footage: behaviors missing from both the human and robot recordings saw no comparable benefit.